Cortical Steganography A Novel Approach

Cortical Steganography:

A Novel Approach to Multifactor Authentication Through Sensorimotor Coupling ✩

A. Chetri ∗

Faculty of Science & Technology, Department of Psychology, 115 New Cavendish Street, London, UK University of Westminster 1,∗

Abstract This paper offers an efficient solution to the rubber-hose cryptanalysis problem by training participants a high-entropy

underlying target sequence in a visuomotor typing task until the skill is sufficiently acquired. On average, the target sequence reaction time was found to be significantly lower in the trained group when compared to the untrained group. But most important of all, participants failed to recall overall sequence structure, establishing tacit skill expression outside of explicit awareness. Therefore, ensuring that an authentication protocol is built around what users implicitly learn through prior sensorimotor training. Thus guaranteeing that critical information is not susceptible to coercion. The findings hope to lay a framework for future crypto primitives in large-scale security protocols.

Keywords: Implicit memory, multisensory processing, security, cryptography

1. Introduction protocols have incorporated numerous measures and tech- niques ahead of cyber-attackers. Such countermeasures of-

Motivations. From Charles Babbage’s Analytical Engine ten include multifactor authentication methods (biometric to the birth of TCP/IP and the internet, it is patently clear

validation, picture recognition, key codes, security ques- that the way we live our lives in the modern era has be-

tions) as well as strengthening network integrity through come increasingly dependent on computers. Consequently,

firewall, encryption, anti-virus, and security-aware policies this has pushed the collective drive to develop robust com-

(trusted platform modules and control lists). Despite all puter security in order to ensure the safekeeping of pri-

this, the possibility of coercion to reveal critical charac- vate information and preventing malicious intent. This

teristics of the network infrastructure (i.e. releasing pass- can include protecting a variety of information pertaining

words) has been little addressed by security researchers. to banking details, medical records, email exchanges, and

Such flaws are often referred to as Rubber Hose Crypto- other types of sensitive data. However, no security system

analysis, and thus presents a major issue in modern com- is error-free and so discernible flaws are often exploited

puter security (Jain and Nandakumar, 2012; Bojinov et by unsavoury users (known as black-hat hackers) as most

al., 2012). In particular, access authorisation through the notably seen with the rise of hacktivism. This can include

use of password authentication can be viewed as inher- planting malicious software (Malware), distributing mal-

ently flawed in terms of rubber hose crypoanalysis-related ware which masks itself as a safe software (Trojan), denial

attacks (Bojinov et al., 2012) given that computer secu- of service (DoS), eavesdropping, spoofing, tampering, re-

rity is limited to protecting computers, and not humans. pudiation, and/or leaking data among others.

Therefore, applications from cognitive neuroscience and By its very nature, computer networks that are built

psychological research could be used. to connect large infrastructures are often targeted and ex-

ploited ahead of smaller networks. Therefore, vulnerabil- Previous investigations. In terms of cognition, this spe- ities within financial sectors, military departments, con-

cific transfer of information from digital to biological sub- sumer electronics, the aviation industry, government, and

strates is achieved through the memory encoding process, large corporations are kept to a minimum by various se-

where the type of memory is very much reliant on the cir- curity departments. With this in mind, modern security

cumstance of encoding. For this reason, most often when learning passwords (or the sensitive information behind it)

the user attends to critical details of the information in or-

✩ This report is submitted in partial fullment of the requirements

of the University of Westminster for the award of BSc (Hons) Cog-

der to later recall the information to a degree of accuracy.

nitive Neuroscience.

To which, this encourages the user to encode explicit mem-

∗ Corresponding author

ories of the information which is processed through con-

Email address: research@ashchetri.co.uk (A. Chetri) URL: www.ashchetri.co.uk (A. Chetri)

scious effort (Squire, 2009). Within this, the conscious- scious effort (Squire, 2009). Within this, the conscious-

Our current understanding of memory tends to distin- guish between aspects of recognition and recall based on conscious awareness of the particular information that is retrieved (Henke, 2010). For example, if an individual re- calls a particular memory yet fails to recognise it, then it is understood that the initial trace may have been specifically consolidated through an unconscious process. Discussions into non-conscious processing and memory have been dis- cussed by various researchers, where implicit memories are said to form outside of conscious processing (Dutta et al., 2014). Therefore, manipulating implicit memory function- ing may be key to storing sensitive information such as passwords without conscious awareness. Hence, acting as a possible solution to the rubber hose cryptoanalsysis prob- lem (Bojinov et al., 2012; Weinshall and Kirkpatrick, 2004; Denning et al, 2011).

As described by Squire (2009), explicit memory is often referred to as declarative memory, which involves memo- ries relating to facts or experiences. One particular as- pect of explicit memory, is that explicit memories can

be conceptualised into knowledge and therefore verbalised through conscious means. Explicit memory can be tested by recall and recognition, and strictly requires effort and intention. However, explicit memories are not durable and can be susceptible to contamination through paired as- sociative learning. On the other hand, implicit memory, otherwise known as procedural memory, involves the learn- ing of skills, habituation, priming, and classical condition- ing. The characteristic of implicit memory which makes it an ideal candidate for psychobiometric research is that implicit knowledge is inaccessible due to unconscious pro- cessing. Nevertheless, knowledge can be expressed through several measures not yet fully understood in the literature.

Over the years, various researchers have dabbled with implicit learning in the context of computer security. One Microsoft Research (Stubblefield and Simon, 2004) report implemented the use of automatically generated Rorchach ink blots for password synthesis. With each ink blot repre- senting a character used to create a semantic value in the form of a password. However this study makes no effort in inhibiting explicit processing from consciously encoding the passwords, namely through memory techniques such as mnemonics. Participants can therefore divulge informa- tion about the inkblot values under coercion as a result of this protocol. In the same year, Weinshall and Kirkpatrick (2004) studied the nature of behavioural imprinting which allows for gradual learning of rich and complicated experi- ences. The study looked into various paradigms involving picture recognition, objective recognition, and pseudoword recognition used as a method for creating secure crypo- graphic tokens as an adaptive model for securing pass- words. Across all paradigms, information was presented in a seemingly random manner under a common theme. Once rehearsed, authentication was later undertaken to verify if the information has been processed, but also in an

implicit manner. The authentication task involved recog- nising items that were previously seen in the initial set. This is repeated after a number of times to ensure sus- picious behaviour such as guessing is accounted for. Af- ter comparing the results with a simulated imposter, the trained group performed significantly better than the con- trol. Also, further studies revealed high levels of durabil- ity with 90% accuracy after 1-3 months, and later 70-80% after over 2 months. However, the study posed multi- ple flaws such as the saliency of items affecting consoli- dation. For one, object recognition and pseudoword gen- erated produced less than favourable results compared to image recognition. Secondly and most importantly, if two or more items contained shared central themes, then this would often lead to each item being confused as a single trace. Thus, presenting a problem for memory recognition, and in turn, authentication as a whole. In order to address this, Gaddy and Ingram (2014) employed a partial picture recognition task, which was first introduced by Srinivas et al (1991).

Fragmented images are used with item familiarisation being the ideal measure in place of total recognition. The research found no significant result in the data to suggest this method could be a viable security protocol. Neverthe- less, the study revealed that encoding implicit memories was highly prone to emotion, and fatigue. Since the stim- uli was lengthy to authenticate, not only would this pose

a problem for practical use, but this also fed into user fatigue and anxiety, affecting consolidation Gaddy and In- gram (2014). Denning et al (2011) also presented a similar study juxtaposing fragmented images with their completed counterparts for familiarization based on image fidelity. However, the results also produced unfavourable results much like Gaddy and Ingram (2014). Moreover, Rossion and Pourtois (2001) suggested that the use of image and object recognition for such studies are troublesome since central themes and motifs are attended consciously and thus, transferred to explicit memory stores. This leaves a substantial wealth of partial explicit information which a potential interrogator, if successful in coercion, may use to piece together the full information. Kulkarni et al. (2012) also comments on the flaws of graphical recogni- tion by suggesting it creates a pertinent conflict between protocol usability and overall security. Therefore, another method of implicit encoding in password security must be explored. According to Henke (2010), implicit memory re- lies upon a merged unitised representation, therefore indi- vidual reactivation of cues cannot suffice. Instead, implicit memories is reliant upon an element of sensory coupling during the learning episode in order to fully encode into implicit stores. Taking this into account, Bojinov et al (2012) sought to explore the area of sensorimotor integra- tion in implicit encoding.

In Bojinov et al (2012)s study, participants implicitly learnt statistically predictable bigramic sequences through

a motor-based Serial Interception Learning (SIL) task. Dur- ing the study, participants expressed improved performance a motor-based Serial Interception Learning (SIL) task. Dur- ing the study, participants expressed improved performance

Figure 1: SIL visuomotor task comprised of 6 keys (SDFJKL).

Theoretical Framework. Nondeclarative memory As mentioned by Henke

(2010), long term memory can be viewed as a hierarchy based upon processing. The main divisions are factored by conscious (declarative memory) and unconscious (non- declarative memory) recognition systems. Contained within the declarative memory branch are episodic (periodic life events) and semantic (facts and general knowledge) mem- ory stores. Neuroimaging evidence has shown activation of both episodic and semantic memory expression in the medial temporal lobe and diencephalon. On the other hand, the nondeclarative memory branch consists of 4 dis- crete stores; procedural or implicit memory (skills and habits), priming (the altering of perception through pro- cessing specific stimuli), simple classical conditioning (as- sociative learning), and habituation (sensitisation) stores. Also neuroimaging data shows procedural and implicit mem- ory activation in the Basal Ganglia, while priming is found in the Neocortex, and simple classical conditioning in the Amygdala and Cerebellum. Although habituation and sensitisation, two aspects crucial to procedural and im- plicit memory is mostly associated with various reflex path- ways. Since rubber hose cryptanalysis entails explicit (declar- ative) processing to be the root cause of the problem, non- declarative stores or namely implicit and procedural func- tioning must be looked at.

The discovery of implicit memory was made by War- rington and Weiskrantz (Warrington and Weiskrantz, 1974 cited in Mancia, 2006) through studies into priming with

Figure 2: The hierarchy of Long-Term Memory according to their processing characteristics via neuroimaging (Henke, 2010).

Korsakovs amnesia sufferers who exhibited damage in ex- plicit (conscious, declarative) stores, yet showed signs of a intact implicit memory function. Further to this, sensori- motor skill learning, priming, conditioning, as well as cog- nitive skill learning have all been found to be spared in am- nesic sufferers, confirming Mancia (2006)s ndings (Henke, 2010). In addition to this, the spared nondeclarative stores have also shown signs of functional perceptual learning, vi- sual categorisation, prototype learning, and tactual maze learning (Henke, 2010). Along with this, artificial gram- mar learning was also intact suggesting that cognitive skill learning is preserved (Knowlton, Ramus and Squire, 1992). All of this points to the sheer robustness of nondeclarative stores, which share an ideal characteristic of unconscious processing requirements. Therefore coupling nondeclara- tive functioning with sensorimotor integration as described in Bojinov et al (2012) proves to be an ideal method for this area of study.

Figure 3: The executive structure of the information processing con- trol system.

The motor system Most studies on implicit mem- ory extend upon the theoretical understanding of motor learning, which cognitive research has largely ignored in most memory-processing models (Whiting, 1989) hence why this has been referred to as the Cinderella of Psy- chology (Rosenbaum, 2005). Nevertheless, studies have pointed towards the link between tacit comprehension and improved motor performance after systematic practice (Mag- ill, 1998). Moreover, there has been suggestions of a rela- tionship between implicit learning and motor action where the greater the implicit elements learnt, the greater the motor performance (Wulf and Schmidt, 1997). On top The motor system Most studies on implicit mem- ory extend upon the theoretical understanding of motor learning, which cognitive research has largely ignored in most memory-processing models (Whiting, 1989) hence why this has been referred to as the Cinderella of Psy- chology (Rosenbaum, 2005). Nevertheless, studies have pointed towards the link between tacit comprehension and improved motor performance after systematic practice (Mag- ill, 1998). Moreover, there has been suggestions of a rela- tionship between implicit learning and motor action where the greater the implicit elements learnt, the greater the motor performance (Wulf and Schmidt, 1997). On top

Figure 4: The expanded structure of the information processing con- trol system, including the executive and effector.

Further to this, the control system models have been extended to account for full motor performance in order to conceptualise the connection between the processing of sensory inputs into muscular output. The expanded model proposes that the executive acts as a mediator in which a directionality state is maintained in order to process into the effector (Schmidt and Wrisberg, 2008). The effector is composed of the motor program, the spinal cord, and muscles necessary to distribute and produce a desired ac- tion (Bapi, Doya and Harner, 2000). This model is then classified into either a closed or open loop control systems based upon either the use of feedback, error detection, correctional process for fine motor control (closed-loop) or no feedback for rapid discrete motor action (open-loop). At the heart of both models is the motor program which dictates the action of acquired skill at the executive level through an array of motor instructions (Schmidt and Wris- berg, 2008). In Bojinov et al (2012)s study, a keyboard was implemented to record keystroke reaction time in order to carry out executive sensory sequence learning facilitated through effector action. Considering this, a closed-loop model of keyboarding may be useful in the overall concep-

tualisation of the motor control contributions to implicit memory function discussed in Bojinov et al (2012)s work.

The visual system Nevertheless, motor function in terms of sensory memory processing is not restricted purely to finger and/or limb movement.

Figure 5: The Optomotor Cycle.

Another important area in this research lies within ocularmo- tor function, specifi- cally saccades, rapid eye jumps (100-300ms) that fixate items of inter- est within the fovea (Rosenbaum, 2010). In

a general sense, sac- cades are conducted in order for the brain to create a rapid account of the visual field through 3-5 sac- cadic snap shots (Rosenbaum, 2010). This includes the in- termediatary events and breaks between eye-movements, suggesting a use of a temporary memory store. It is there- fore understandable to suggest that the fixation allows in- dividuals to attract spatial attentional resources into the fixated image. The link between spatial attention and sac- cades was first suggested by Posner (1978) who proposed that reaction time is lowered when spatial (and movement) information of objects are anticipated to a degree of accu- racy. Although, recent evidence has suggested that it is conversely implicit learning which guides attentional re- sources via saccades (Jiang, Won and Swallow, 2014).

According to the optomotor cycle, conscious attention mediates the stop and go action of the oculomotor sys- tem. In this model, the fixation acts as a saccadic suppres- sor which stops eye movement, whereas the reflex gener- ates saccades and promotes eye movement (Fischer, 2007). The optomotor cycle between saccades and antisaccades, governed by attention, can be seen to guide the learning process and enhance plasticity in infants during early de- velopment (Yu and Smith, 2011; Fischer, 2007). In this way, it should not be surprising that as the more dominant of the other sensory systems, the role of visual systems is also paramount in motor coordination, which in this case would be the finger pressing of keystrokes. The literature tends to be fairly heavy on how vision underpins moment and fundamentally human behaviour, with two particular systems for object identification (focal and ambient vision) derived from the understanding of two distinct visual path- ways; the ventral and dorsal streams (Goodale and Mil- ner, 1992; Trevarthen, 1968). According to this 2-stream hypothesis (Mishkin and Ungerleider, 1982; Goodale and Milner, 1992), as visual information exits the occipital lobe (the brain region chiefly involved in vision) the sensory information is distributed into two pathways based upon object identification in the temporal lobe (ventral, what pathway) and spatial location in the parietal lobe (dor- According to the optomotor cycle, conscious attention mediates the stop and go action of the oculomotor sys- tem. In this model, the fixation acts as a saccadic suppres- sor which stops eye movement, whereas the reflex gener- ates saccades and promotes eye movement (Fischer, 2007). The optomotor cycle between saccades and antisaccades, governed by attention, can be seen to guide the learning process and enhance plasticity in infants during early de- velopment (Yu and Smith, 2011; Fischer, 2007). In this way, it should not be surprising that as the more dominant of the other sensory systems, the role of visual systems is also paramount in motor coordination, which in this case would be the finger pressing of keystrokes. The literature tends to be fairly heavy on how vision underpins moment and fundamentally human behaviour, with two particular systems for object identification (focal and ambient vision) derived from the understanding of two distinct visual path- ways; the ventral and dorsal streams (Goodale and Mil- ner, 1992; Trevarthen, 1968). According to this 2-stream hypothesis (Mishkin and Ungerleider, 1982; Goodale and Milner, 1992), as visual information exits the occipital lobe (the brain region chiefly involved in vision) the sensory information is distributed into two pathways based upon object identification in the temporal lobe (ventral, what pathway) and spatial location in the parietal lobe (dor-

Figure 6: Neuroimaging data shows patterns of activation in areas implicated in implicit memory, sensorymotor, and visual processing (Moussa et al., 2012).

Therefore in a broad sense, the visual system can be viewed in terms of stimulus identification-recognition and visually guided behaviour. Considering this model, for efficient implicit sensorimotor learning, small amounts of conscious visual information may be required in order for the encoding process to take place effectively.

Although not widely accepted, the Default Mode net- work (made up of the precuneus/posterior cingulate, infe- rior parietal lobes, and medial frontal gyrus) in the brain is understood to be active during low-attentional focus and wakeful rest (Broyd et al., 2009). As measured through fMRI resting state network (RSN), the Default Mode is found to be involved with the attentional focus switch from resting state to high focus (Fox et al., 2005). This was also confirmed in later neuroimaging data, where the visual network’s capacity for RSN attentional switch was found to be highly consistent with sensory-motor, and basal gan- glia (implicated in implicit learning) cortices; but not so much for the Default Mode network (Moussa et al., 2012). Other regions such as the lateral and ventral intrapari- etal cortex of the intra-parietal sulcus (IPS) in the parietal lobe is understood to play a significant role in visual atten- tion and saccadic eye movement (Culham and Kanwisher, 2001). The IPS is also strongly implicated in sequence learning through fine finger moments through perceptual motor coordination (Sakai et al,. 2002). Such findings uni- fies several highly functional visuomotor networks and at- tention with the idea of a Visual Association Cortex (Zeki, 1993).

Keeping this in mind, the individual and combined roles of the two specialised visual systems, focal vision and ambient vision, is also fundamental to the visuomotor

structure described. Focal vision specialises in the con- scious visual identification of the stimuli for an appropri- ate input response as described in the information process- ing model. Through the use of attention, the focal visual system allows individuals to consciously fixate to form a perception of objects of interest (Schmidt and Wrisberg, 2008). Ambient vision on the other hand is largely respon- sible for movement control in both central and peripheral visual fields. Unlike focal vision, ambient vision detect the spatial location and moment of objects in relation to the individual. But most importantly, the ambient visual sys- tem contributes largely to fine-motor control without con- scious awareness (Schmidt and Wrisberg, 2008). The way in which focal attention orientates attention into muscular output is based upon feedback and error correction, which contributes to calibrating attention to information on the surroundings for efficient motor performance (Schmidt and Wrisberg, 2008).

Attention In order to inhibit conscious awareness of the trained item(s), an authentication protocol based upon implicit learning must discriminate between explicit and implicit processing. It is therefore worth investigating the role of attention as a mediator of information process- ing for this particular area of study. To which, Lozito and Mulligan (2010) looked into the role of attention in the implicit memory retrieval process during test phase. Within the study was three separate experiments based upon perceptual identification, word-stem completion test, and category exemplar production test. Out of all the tasks, attention was found to strengthen the performance of a secondary task through ’automatic’ retrieval.

However, there are several types of attention to con- sider depending on capacity and selectivity (Wickens and McCarley, 2008). Although, one view is that attention can either be classified as a capacity or a resource, where in- terference can occur from the conflict between attentional capacity and available resources for allocation (Schmidt and Lee, 2011). Interference often manifests as disrupted performance, namely in the speed of processing which en- tails a disruption of the autonomic (non-conscious) ele- ments of attention (Schmidt and Lee, 2011). As previ- ously discussed, it is clear that attention and conscious- ness are closely linked in tandem as first speculated by James (1890). But also in modern years, studies involving performance-dissociation tests revealed distinct behavioural contributions between conscious and unconscious atten- tion (Jacoby, Lindsay and Toth, 1992). Conversely, at- tention can also be conceptualised as either a controlled (conscious) or automatic system (unconscious) to address theoretical issues with performance errors referred to as action slips (Norman, 1981). One aspect not covered yet is attention’s tendency to be highly susceptible to emo- tional arousal. For example, anxiety and stress have been hypothesised to restrict performance (hyper-vigilance) due to a decrease in cue utilisation which may cause a break- However, there are several types of attention to con- sider depending on capacity and selectivity (Wickens and McCarley, 2008). Although, one view is that attention can either be classified as a capacity or a resource, where in- terference can occur from the conflict between attentional capacity and available resources for allocation (Schmidt and Lee, 2011). Interference often manifests as disrupted performance, namely in the speed of processing which en- tails a disruption of the autonomic (non-conscious) ele- ments of attention (Schmidt and Lee, 2011). As previ- ously discussed, it is clear that attention and conscious- ness are closely linked in tandem as first speculated by James (1890). But also in modern years, studies involving performance-dissociation tests revealed distinct behavioural contributions between conscious and unconscious atten- tion (Jacoby, Lindsay and Toth, 1992). Conversely, at- tention can also be conceptualised as either a controlled (conscious) or automatic system (unconscious) to address theoretical issues with performance errors referred to as action slips (Norman, 1981). One aspect not covered yet is attention’s tendency to be highly susceptible to emo- tional arousal. For example, anxiety and stress have been hypothesised to restrict performance (hyper-vigilance) due to a decrease in cue utilisation which may cause a break-

encoding process of implicit memory is also highly vul- nerable to emotion (Mancia, 2006). More still, attentional interference also disrupts familiarity and recollection in im- plicit memory recall (Topolinski, 2012). This is very much akin to how interference disrupts automaticy, unconscious skill expression (Schmidt and Lee, 2011). Which in turn, feeds into motor performance through the measure of re- action time (Bojinov et al., 2012). As discussed, the motor

Figure 7: Specific mapping (A) of fingers to 6 keys (SDF, JKL)

system is a trigramic relationship between the task, the in-

in Bojinov et al. (2012) compared to novel method of non-specific

dividual, and the environment (Repp and Su, 2013). How

mapping of 6 keys within determined areas (B).

this feeds into perceptual learning is another dimension can be explained through sensorimotor skill expression as

can be demonstrated in the expression below.

a function of implicit (procedural) memory (Henke, 2010). Although, enhanced motor learning is highly associated

Lef t Hand = {x 1 ,x 2 ,x 3 }, Right Hand = {y 1 ,y 2 ,y 3 } with systematic practice and habituation which results

To verify the usability of such a wide space of keys, a num- in permanent and autonomous (unconscious) performance ber of measures will be taken. Namely, analysing the com- (Schmidt and Lee, 2011). Combining these lines of evi- bining properties of how the 32 keys is spread across two dence; attention, visuomotor system, perceptual learning, spaces, and the chances of an unauthorised user guessing and implicit memory all share substantial commonground

a correct combination. Given that keys 1-16 are within for further research.

a set of the left keyboard space (x) and keys 17-32 are As with Bojinov et al (2012)s findings, repeating ele- within a set of the right keyboard space (y), the number ments of a structure were superior to the implicit learn- of possible permutations of all 16 keys for each keyboard ing than the non-repeating (and sometimes changing) el- space is 313,600, as expressed below. ements (Pew, 1974; Magill, Schoenfelder-Zohdi, and Hall cited in Maghill, 1998; Wulf and Schmidt, 1997). With the 2 16!

pertinent problem presented in rubber hose cryptoanalysis

3 3 3!(16 − 3)! (and the findings from previous implicit learning studies

pointing less favourly to visual recall) the method of mo- This greatly improves upon Bojinov et al (2012)s keyboard tor learning to aid implicit learning of motor commands 2 layout as there can only be (4C3) = 16 possible combi-

to produce a signature performance may be a viable area nations in the previous study. Therefore, the layout used to explore in security research. However, Bojinov et al

in this study is an effective solution for non-authorised (2012)s design could be improved upon, alongside inves-

users pre-empting keystroke patterns due to humans form- tigating effects of saccaddic manipulation and the role of

ing a symbolic internal layout for motor planning effects attention in the encoding process. Although in order for

(Krakauer and Shadmehr, 2006). This creates a robust the protocol to act as a viable security system, efforts must

platform for developing strong and secure sequence syn-

be taken to lower the authentication time. All the while,

thesis.

expression of skill must be viewed as the password rather than the item being rehearsed, as ultimately authenitca- tion will identify correct users based upon performance. Overall, this creates a strong situation-based approach by combining context (environment), sensory-perceptual de- mands (task), and motor performance (individual) into a framework for the rubber hose cryptoanalysis problem.

2. Methodology

Figure 8: Euler cycle was used to generate 30 bigrams keys.

Key Selection & Cluster Topology. An array of 32 key- board keys discriminated by left hand (x keys) and right

In the same manner as Bojinov et al (2012), using the hand (y keys) was uniformly distributed into paired bi-

bigram matrix generated from the 6-node Euler circuit, grams sequences for optimal motor performance and en-

the 30-character alphanumeric Target sequence (T) was hanced security.

synthesised as expressed below.

For this, 3 keys per hand are selected on a pseudo- random basis using a software number generator, and thus

Sequence Length : {x 1 ,x 2 ,x 3 ,y 1 ,y 2 ,y 3 } = 30

To synthesise the primary sequence, T, 15 bigrams from the matrix are selected at random from the pool, creating a 30 character string. Whereas for the secondary sequence, the Foil (F), 9 bigrams are selected at random from either the remaining pool of bigrams or the entire matrix, depending on cluster typing. A Cluster is made up of the 2 sequences (Target and Foil), with the Target repeating 3 times and the foil repeating once. Therefore a typical Cluster for this study can be expressed as:

3T + F

Moreover, the distribution of the bigrams in the com- position of T and F is also dependant on cluster typing. Cluster A and Cluster C will be randomly ordered, Cluster

B will have a determined order to ensure less overt pat- terns occurring in the sequences, but also patterns which can be construed by the user to make semantic sense (e.g. HI, LO = High Low).

The method of synthesising an 18-character secondary sequence, F, from the remainder of bigrams is also depen- dant on cluster typing. The property of Cluster A is such that F is synthesised from the entire alphanumeric pool, which includes items that have already been assigned to a sequence. This creates a less-salient overlap between Tar- get and Foil sequences. Whereas Cluster B generates F by randomly selecting 9 bigrams from the remaining pool, creating a more salient relationship between T and F. In Cluster C, elements of the secondary sequence is derived from a completely new matrix, and so contains a higher number of keys used (6 keys per hand). This creates a far more salient perception of T and F, while also, disrupting the motor planning effects of T when shifting to a com- pletely new layout in the F sequence. However because of the clear separation between the sequences, participants may perceive an underlying structure inhibiting implicit encoding of T. Hence why cluster typing analysis is in- cluded in the investigation.

Experimental design. Participants will be arranged in two groups in an independent groups design. The first group will be categorised as the trained group, who will receive prior training before undergoing a test phase. The second group will be the untrained group, who will only partake in the test phase without prior training. The research ques- tion aims to address whether training of target sequences has an effect on target sequence-specific performance rates via reaction time.

In order to investigate the effects of sensorimotor pre- training on test phase reaction time, reaction time will

be measured as the continuous dependent variable. Al- though percentage advantage for specific sequences over the other will also be measured as a modulus of reac- tion time. Whereas, the training vs no training category, and the session category being the between-groups fac- tors. Within-groups factors such as sequence type (Target or Foil) performance will also be taken into account.

The first hypotheses are two-tailed and will concern the sequence-level investigation, with the null hypothesis being that mean reaction time between Target and Foil sequences are equal for both groups. Whereas the alterna- tive hypothesis in this level is that the mean reaction time difference between Target and Foil sequences are unequal for both groups.

H 0 : target = f oil, H A : target 6= f oil Finally, the second hypotheses will concern the one

tailed group-level investigation, with the null hypothesis as the mean reaction time of target sequences in the trained group will be equal to the mean reaction time of target sequences in the untrained group. On the other hand, the alternative hypothesis will be that mean reaction time of target sequences in the trained group will not be equal to the mean reaction time of target sequences in the un- trained group.

H 0 : trained = untrained, H A : trained 6= untrained Participants.

2 groups of 5 healthy adult participants (n =

10) were voluntarily recruited for the study through Uni- versity mailing lists and noticeboard advertisements. 40% of the participants were male and 60% were female, with the youngest participant being 18 years old and the old- est being 55 years old. One participant withdrew from the study, and therefore was not included in the final data pool. Therefore, the participant’s control counterpart was also not recruited as no comparisons could be made oth- erwise.

Apparatus. The software (KeyMapper) for generating bi- grams was developed for the study. To build the pro- gram, the code was written in Swift (programming lan- guage) via XCode (XCode, 2015). KeyMapper randomly assigns 3 keys per hand for both left-right hand keyboard spaces, and randomly generates 30 bigrams from the keys produced. The application is compiled for a Mac OS X environment, however the methods used can be repro- duced manually on paper. As well as KeyMapper, the stimulus (CorticoSteg) used in the experiment was pro- grammed prior to the study using Game Maker software (Overmars et al., 1999) on a Mac OS X (UNIX) oper- ating system environment. Therefore, the final software build is a Mac OS X application written in Game Maker Language (GML). CorticoSteg runs the trials by display- ing the stimulus for the participants, recording the times- tamp for each keystroke in a .csv file format, and making

a distinction between each incorrect, missed, and correct keystroke. Finally, a standard QWERTY keyboard was used throughout the onset of the study. In order to con- trol for minute differences between QWERTY layout de- sign, the same keyboard was used.

Procedure. Despite sequences being arranged in bigrams, the bigrams will be grouped in order to appear as paired- trigrams in the final stimuli (e.g. bigrams FI, P8, UJ -¿ FIP, 8UJ). A trigram of consecutive sequence items shift- ing from box 1 to box 2 to box 3 one at a time. This

method was selected due to saccadic eye movements play- ing a significant role in the spatial anticipatory factor of oculomotor sequence learning and its effects for proce- dural memory consolidation (Albouy et al., 2008). But more importantly, trigrams facilitates the user to antici- pate motor patterns in groups of 3, which in turn allows for stronger performance advantage (Destrebecqz and Cleere- mans, 2001).

Figure 9: An example of the difference between the tasks. Unlike the 35-trial training phase, the 3-trial test phase involves the participant to quickly type keys including hidden elements of the sequence with no time constraint.

Primarily, the experiment will be carried out in two phases, the training phase and the test phase. The train- ing phase will involve participants typing keys that quickly appear on the screen in a seemingly random manner. Keys displayed will be ordinal, dependant on cluster topogra- phy. So therefore, the participant will be typing both Target and Foil sequences. However, since the Target se- quence is repeated 3 times and the Foil once, the partici- pant should be better trained on the Target sequence. It is understood that this comprehension in Target sequences over Foil sequences should manifest in motor performance over the course of the training phase. And so, such per- formance should involve faster reaction time and fewer in- correct keystrokes for the Target sequences compared to Foil sequences. Each cluster will be considered as 1 trial, with 5 trials per block. Thus in total, the training phase will comprise of 7 blocks (35 clusters, 3780 items) with a maximum timer of 2 seconds per item. This is enforced to ensure that items that exceed the temporal window will

be skipped and considered a miss (incorrect keystroke), in- hibiting the user to attend to the item promoting explicit processing. Therefore, the issue of explicit memorisation techniques such as chunking can be addressed through the sheer speed of the task. Participants will be given a break upon competition of each block.

Within the test phase, there will be 2 further tests sep- arated by at least 24 hours (Session 1 and Session 2). This phase is almost identical to the training phase, however, there is no timeout and various elements of the sequences

will be hidden. Therefore the test phase will rely on both the speed, and correct keypresses which will be the hall- marks that learning has taken place. The second session is purely to test the robustness of the learning through observing retention rates (performance over the course of

2 sessions). A final verification of the type of learning will

be made by asking the participant to explicitly reveal the sequence(s).

Participants will be divided into 2 groups, the Trained group and the Untrained group. The Trained group will undergo the training phase prior to the test phase, while the Untrained group will only undergo the test phase (al- beit, with explicit knowledge of the sequence for Session 1). Cluster typing for this study will be predetermined prior to the study, with 2 participants in each group be- ing assigned a cluster type. Therefore, participants 1-2 (group 1) and 6-7 (group 2) will be assigned to cluster A, while participants 3-4 (group 1) and 8-9 (group 2) will be assigned to type A. Due to the attrition of 1 participant, Cluster C will not be assigned to 2 participants per group. The choice to not recruit the group 2 Cluster type C will

be assigned to the remaining participants, participant 5 (group 1) participant 10 (group 2).

Ethics. In terms of ethical implications, the study aimed to keep participant harm at a minimal level. However, there may be acute issues that may have arisen during the course of the study. Namely, there may be issues with participant eye-strain during lengthy training phase ses- sions. In a broad sense, the average duration of each

training phase was around 40-50 mins (with breaks be- tween blocks). Despite this, the stimuli was coded on a white background on black text. But since the study re- quires the participant to make saccadic eye movements across each box, this may harm the participant.

To address this, a screen tint was offered to each par- ticipant at the beginning of the session. However, no par- ticipant made use of this. It could be the case that the participant may have been naive to the extent of how the stimuli will affect their eyes prior to the start of the study. Although, allowing the participant to use a screen tint mid-way may affect results and thus poses a flaw in the methodology in maintaining controlled conditions of the experiment. With this in mind, a grey background will be used for all participants in future studies.

3. Results Training Phase.

Sequence performance With the learning curve (fig- ure 10), reaction time performance between target and foil sequences was compared. Through a one sample t- test, Target sequence reaction time was demonstrated to show significantly higher knowledge advantage than Foil sequence reaction time (t(34) = 8.66, p < .05). Block 1 (7.17%) to Block 2 (21.76%) provided for a highest leap

Training phase model Through the data collected from overall sequence-level performance, a logarithmic re- gressional model for both Target and Foil sequences for each cluster was made (figure12). From this, predictions for future performance for 2 additional blocks were fore- casted. Type A and Type B clusters display similar log-

arithmic patterns, with a minor 0.11265 R 2 -difference in Foil sequences. Additionally, there is also a much closer ±0.08761 difference in Target Sequences. However, as pre- dicted, Type C clusters show less Target-Foil disparity. As such, there is not enough sufficient data found in Type C clusters in order to create a robust model on the sequence interactions observed in Type A and B. However, Type C cluster performance is still important for comparative rea- sons. The issue of structural saliency (typical to Type C clusters) affecting Target-Foil reaction time disparity can

Figure 10: Target sequence disparity rates over foil.

be applied when formulating effective sequence structures for optimal performance.

(c) Type-C (a)

(a) Type-A

(b) Type-B

(b)

Figure 12: Cluster-level regressional analysis through logarithmic Figure 11: Participant-level data for mean reaction time (a), and the

models to predict future block performance. target sequence advantage rates (b)

in the performance advantage of Target sequence reaction time, compared to and Foil sequence reaction time. In Block 2 to Block 5, reaction time stabilised around the 21% mark in performance advantage. However, this soon dips slightly in Block 5 (21.47%) to Block 6 (18.36%) before rising further in Block 7 to make for the second highest in- crease (28.88%). In addition, participant-level differences (figure 11) was also extracted from the data with partic- ipant 1-2 (cluster A), participant 3-4 (cluster B), partici- pant 5 (cluster C). On average, there is a 22.96% difference in Target sequence-specific knowledge advantage over Foil sequences. Participant 5 had the fastest reaction time, however displayed a minuscule Target-specific advantage over Foil reaction times (0.17%). Participant 1 had the fastest reaction time with a 3rd highest magnitude of dif- ference between Target and Foil sequences (24.04%).Par- ticipant 4 also exhibited the largest reaction time differ-

Figure 13: Overall cluster-level regressional analysis

ence between Target and Foil sequences (65.70%). Par- ticipant 2 displayed the second highest predicted diver-

Therefore, a viable security protocol based upon motor gence (25.67%). Finally, Participant 3 showed a mod-

knowledge recall must include a degree of non-saliency be- est difference between Target and Foil sequence reaction

tween Target and Foil sequences. An overall model includ- times (9.41%), the second lowest Target-Foil predicted di-

ing all cluster types was also synthesised from the data, vergence.

with both models including a negative natural log coef- ficient, suggesting an increase in keystroke speed through lower reaction time (figure 13). This fits in the overall pre- with both models including a negative natural log coef- ficient, suggesting an increase in keystroke speed through lower reaction time (figure 13). This fits in the overall pre-

arated by approximately ±0.1 difference between Target edge of the box. These values were not omitted from the and Foil sequences, however the reaction time disparity is

data as the values did not considerably affect the over- evident through the wide reaction time margin as noted

all outcome of the findings. There was homogeneity of in (Fig 1) and in the intercept difference showing approxi-

variances, as assessed by Levene’s test for equality of vari- mately ±0.1 seconds, a ≈ 100ms difference. Although rela-

ances (p > .05). Mauchly’s test of sphericity was assumed tively small, such a difference can provide a framework for

to be violated as the outcome did not return a significant

a baseline level for crypto primitives when recording per- result. Therefore, a Greenhouse-Geisser correction was ap- formance. In terms of designing these security algorithms,

plied instead. Through this, there was a statistically sig- the data presented above may prove to be especially use-

nificant three-way interaction between sequence, session ful. However there may be issues with a saturation point 2 and group, F (1, 8) = 5.736, p = .044, partial η = .418,

in the separation of explicit and implicit processing. For

example, identifying the point in which over-rehearsal of An independent t-test was carried out to analyse sequence- tacit information feeds into explicit processing. Neverthe-

level differences. During Session 1, the Trained group less, it is important to view such findings as early work to

showed a significant disparity with Target and Foil se- create a framework for further investigation. Overall, the

quences (t(297.853) = −2.246, p = .025). However, the general models presented are as follows:

Untrained group displayed an insignificant difference (t(228.709) =

2 −1.263, p = .208). In Session 2, the Trained group se-

F oil : y = −0.029ln(x) + 0.7662, R = 0.70076 quence disparity was also significant (t(194.930) = −2.824, p =

T arget : y = −0.065ln(x) + 0.6905, R = 0.81377 005), while the Untrained group sequence difference re- mained insignificant (t(478) = −.727, p = .468). Over-

Test Phase. Test phase was evaluated per 3 cycles (3 repe- all across both sessions, the Trained group exhibited a titions of the cluster), with the mean reaction time of each

significant difference between Target and Foil sequences target and foil sequences for every cycle being recorded for

(t(458.145) = −3.608, p = .0001), as the Untrained group later analysis (figure 14). As the calibratory period lies pri-

exhibited an insignificant difference (t(507.081) = −1.443, p = marily within the first cycle, this tends to skew the data

. 150). Therefore, there was a statistically significant dif- greatly to which it is no longer representative of the overall

ference between means in the Trained group (p < .05) and trend. For this, the first cycle for each session was viewed

an insignificant difference between the means in the Un- as a practice phase and omitted from the final analysis.

trained group (p > .05). Hence, the null hypothesis is Therefore, the mean test phase reaction time is adjusted

rejected in order to accept the alternative hypothesis of to reflect a more accurate total measure for the average

the first supposition.

reaction time between Cycles 2, 3 (session 1 and session

Table 1: Overall test phase performance over correct keystrokes.

Mean RT (s)

4.727 In order to measure group-level differences (table 1),

an independent samples t-test was run. The first session saw the trained group perform on average significantly faster on Target (t(489.994) = −3.354, p = .001) but in- significantly faster on Foil (t(232.476) = −1.893, p = .06)

Figure 14: Observable non-parity in Target-Foil is inverse for all

sequences compared to the untrained group. Moreover,

calibratory cycles.

in the second session, the trained group performed on average significantly faster on the Target (t(371.571) =