26 students from the University of Birmingham (9 males; mean age = 19.0 years [SD = 0.9 years, range = 18 – 22 years]) took part in exchange for £16 or course credit after providing informed consent. All procedures were approved by the University of Birmingham Research Ethics Committee (ERN_17-1673) and were performed in accordance with their guidelines. Data from three participants were excluded from all analyses; two performed below chance (d’ < 0), indicating failure to follow instructions, and one withdrew early, leaving 23 complete data sets.
Apparatus & Materials
A DELL XPS-15 laptop PC (screen resolution: 3840 x 2160; refresh rate: 60 Hz) using custom software created in Matlab (The MathWorks, Inc) and Psychophysics toolbox 36 determined stimulus presentation order, recorded data, and presented stimuli. Authentication responses were recorded using an external standard QWERTY keyboard.
All banknotes were presented as digital video images. 48 different genuine banknotes (24 x £20 and 24 x £50 notes) and 24 different counterfeit banknotes (12 x £20 and 12 x £50 notes) were used as stimuli. Genuine notes were sourced from local commercial banks and showed typical signs of prior use. Counterfeit notes, obtained from the Bank of England, had been recovered from general circulation and showed a comparable level of wear. The primary obvious security features were a Recolour Hologram thread (£20) and green micro-optic security thread (£50) as shown in see Figure 6a.
To prepare stimuli for presentation, each banknote was videoed in a well-lit room with natural light sources in front of the banknotes using a Sony FDR AX33BDI camcorder, recorded in 3840 x 2160 resolution (ultra high definition). During filming, each banknote was centred in a note frame attached to a robotic arm that rotated slowly around a vertical axis. This apparatus was positioned in front of a monochrome textured board. Each video was then edited into a 300ms clip for use in the study. For £20’s, video clips were edited to include the security feature hologram’s image to visibly change between “£” and the number “20” and for a bright reflection from the shiny strip to be visible. For £50’s, editing included visible up-down motion in the holographic stripe and no obstruction by light glare. Half of the videos showed a note rotating left-to-right; remaining videos showed a note rotating in the opposite direction. Clips were then re-sized to match the physical size of each banknote. During testing, each clip was presented in the centre of the black laptop LCD screen framed by a grey (RGB: 127 127 127) rectangle. The participant viewed the screen from approximately 50 cm.
Participants were handed a genuine (physical) £20 then £50 banknote (order counterbalanced across participants) and asked to familiarise themselves with each note for one minute in preparation for the upcoming authentication task. Next, they viewed a genuine £20 and £50 banknote video clip on the screen for 3 s as it would be presented in the upcoming authentication task. These presentations were repeated at least five times and up to 10 times if requested, allowing participants to became familiar with the appearance of genuine notes. After familiarisation, the authentication task was conducted. On each trial, the participant viewed a bright, grey rectangle for a jittered interval (average 2 s), followed by a 3 s banknote video clip. See Figure 6b. At offset, the participant reported whether the banknote was counterfeit or genuine by pressing the ‘z’ or ‘m’ key with their left and right index finger, respectively. The next trial began immediately after the response.
The authenticity task comprised 20 blocks of 36 trials each (totalling 720 trials). Half of blocks presented only £20s, remaining blocks presented only £50s. Blocks alternated denomination with order being counterbalanced across participants. 20% of trials within each block presented counterfeit; order of banknote type within each block was individually pseudo-randomised. Before starting the experimental trials, six practice trials were provided using the same procedure as in the main experimental blocks except that feedback (beep for correct responses) was provided for practice trials only.
EEG Recording. Electroencephalographic (EEG) data was collected during the authentication test. The EEG was recoded using active Ag-AgCl electrodes (BioSemi) from 32 scalp sites (FP1, FP2, AF3, AF4, F7, F8, F3, F4, FC5, FC6, FC1, FC2, T7, T8, CP5, CP6, CP1, CP2, C3, C4, P7, P8, P3, P4, PO3, PO4, O1, O2, Fz, Cz, Pz, Oz) and the left and right mastoids according to the 10-20 system (American Electroencephalographic Society, 1994). For the detection of eye movements and blinks, vertical and horizontal electro-oculogram (EOG) was recorded from electrodes placed above and below the right eye and at the outer canthi of each eye. The EEG and EOG were low-pass filtered with a fifth-order sinc filter (half-power cutoff at 128 Hz) and digitized at 512 Hz.
ERP Processing. All ERP data analysis was conducted using EEGLAB 37 and ERPLAB Matlab toolboxes38, as follows. The EEG signals were offline referenced to the average of the left and right mastoids. The EEG was bandpass filtered offline using a Butterworth infinite impulse response filter with half-power cutoffs at 0.05 and 30 Hz and a roll-off of 12 dB/octave. The data was then down-sampled to 256 Hz.
Noisy channels were substituted by interpolating neighbouring electrode sites. Then, independent component analysis was used to estimate and remove eye-blinks and eye-movements from the stimulus presentation period of the trial using EyeCatch39. Finally, trials were removed if the EEG exceeded ±100 µV in any channel between 200 ms before and 1200 ms post banknote onset: these included noisy segments (eye-blinks, movement, muscle tensing, etc…). Trials were also excluded if the vertical EOG exceeded ±80 µV between 200 ms prior to and 200 ms post banknote onset to ensured that the eyes were not closed when the stimuli were presented.
Averaged event-related potential (ERP) waveforms were computed by averaging trials (200 ms before the onset of the banknote to 1200 ms post onset), after they were baseline corrected to the pre-stimulus interval. The amplitude of the P3 and extended P3 components were measured on the Pz electrode site, as the mean voltage during predefined time windows. The time windows for the P3 and extended P3 components were 450 – 600 ms and 600 – 900 ms post banknote onset, respectively. The P1 was defined as the mean voltage from 50 – 100 ms post banknote onset on the Oz channel.
Multivariate Pattern Analysis. Pre-processing of the EEG data for the MVPA analysis was the same as for the ERP analysis. MVPA was performed on the epoched data (200 ms prior to banknote onset to 1200 ms post onset) using the ADAM toolbox26 in Matlab. A linear discriminant classifier was trained and tested on each time point using a 5-fold cross validation. The area under the curve (AUC) was used to measure classification accuracy. The classifier was first trained on the authenticity of the banknote (genuine versus counterfeit). The classifier was then trained separately for genuine and counterfeit banknotes depending on the response of the participant (genuine versus counterfeit response). Because of unbalanced trial numbers for genuine and counterfeit banknotes, the trial numbers were balanced based on the condition with the lowest trial numbers. Next, temporal generalisation matrices were calculated using cross-classification across time. Statistical analyses for the MVPA were performed using the ADAM toolbox. Cluster-based permutation corrected 2-sided t-tests against chance (0.5) were used to analyse the classification accuracy.
Data Analysis. Three 2 x 2 repeated-measures analyses of variance (ANOVAs) were conducted on ERP magnitude measures using denomination (£20, £50) and authenticity (Genuine, Counterfeit) as within-subject factors. The first ANOVA was conducted on the P1 (50 – 100 ms) component data at electrode Oz, the second on the P3 (450 – 600 ms) component data and the third on the extended P3 (600 – 900 ms) component data, both at electrode Pz.
To analyse the difference between conscious and unconscious neural activity to the banknotes, three further 2 x 2 repeated-measures ANOVAs were conducted on the same ERP components using authenticity (Genuine, Counterfeit) and response (Genuine, Counterfeit) as within-subject factors. The data was collapsed across denomination to increase the number of correct and incorrect trials in the analysis. From these data, only participants who had more than 20 incorrect responses to both genuine and counterfeit notes were included (19 participants).
Counterfeit sensitivity was calculated as d’ for each authentication denomination x fake level condition. d’ was calculated as Z (hit rate) – Z (false alarm rate). Hit rate was calculated as the proportion of counterfeit notes recognised as counterfeit; False alarm rate was calculated as the proportion of genuine notes incorrectly judged as counterfeit. d’ was analysed using a paired-samples t-test comparing denominations (£20, £50). Reaction times (RT) faster than 200ms, slower than 5000ms, and incorrect trials were removed, then RT slower than 3 S.D above the participant x condition mean RT were removed. Remaining RT were analysed using a 2x2 repeated-measures ANOVA with denomination (£20 versus £50) and authenticity (genuine, counterfeit) as within-subject factors was conducted.
All follow-up pairwise comparisons were corrected for multiple comparisons using the False Discovery Rate procedure40. Alpha levels were set at .05 throughout.