Forged Handwriting Detection Hung-Chun Chen M.S. Thesis in Computer Science Advisors: Drs. Cha and Tappert
Motivation Important documents require signatures to verify the identity of the writer Experts are required to differentiate between authentic and forged signatures Important to develop an objective system to identify forged handwriting, or at least to identify those handwritings that are likely to be forged
Key Idea It seems reasonable that successful forgers often forge handwriting shape and size by carefully copying or tracing the authentic handwriting Forensic literature indicates that this is true
Hypotheses Good forgeries – those that retain the shape and size of authentic writing – tend to be written more slowly (carefully) than authentic writing Good forgeries are likely to be wrinklier (less smooth) than authentic handwriting
Methodology Handwriting sample collection Measurement (feature) extraction Speed Wrinkliness Statistical analysis
IBM Thinkpad Transnote
Database Construction Record format for the handwriting samples ID of subject online or offline ID of copied subject word written first/second/third try sampling rate (online) or resolution (offline) file extension
- . <File> Subject ID online offline ID of copied subject xxxx ON OFF yyyy - April T Rate Resolution . Extension Subject ID online offline ID of copied subject word written first try second try third try 100 Hz 300 dpi 600 dpi file extension
Handwriting Samples
Feature Extraction Speed Wrinkliness
Speed The digitizer records the x-y coordinates of the pen movement at a sampling rate of 100Hz This information is used to calculate the average speed of each handwriting sample
Speed The original file of the points ** Page 10 has 4 scribbles: PageSize is 21.59 cm wide by 27.94 cm high. Scribble 0: time 2002/12/11 23:37 Stroke has 93 points: Point ( 4.73 , 5.02 Point ( 4.73 , 5 ) Point ( 4.73 , 4.99 ) Point ( 4.73 , 4.97 ) .... Scribble 1: time 2002/12/11 23:37 Stroke has 113 points: Point ( 5.82 , 5.26 ) Point ( 5.83 , 5.26 ) Point ( 5.85 , 5.25 ) Point ( 5.88 , 5.24 )... Scribble 2: time 2002/12/11 23:37 Stroke has 7 points: Point ( 7.93 , 4.61 ) Point ( 7.94 , 4.61 ) Point ( 7.96 , 4.61 ) Point ( 7.99 , 4.62 )... Scribble 3: time 2002/12/11 23:37 Stroke has 47 points: Point ( 8.26 , 5.75 ) Point ( 8.27 , 5.75 )....
Wrinkliness Wrinkliness = log( high_resolution / low_resolution) / log(2) high_resolution – the number of pixels on the boundary of the high resolution handwriting sample low_resolution – the number of pixels on the boundary of the low resolution handwriting sample Note that the wrinkliness of a straight line = 1.0
Original handwriting sample
Find the edge of the handwriting
Edges of 300 and 600 dpi
Number of pixels on the boundary Convert the scanned images to color images Count the number of pixels whose (Red < 50, Green < 50, Blue < 50) in two different resolutions Get the wrinkliness value
Sample Results Filename 300dpi 600dpi Wrinkliness Speed
Information of the ten subjects UserID Age Ethnicity Education Gender Schooling Handiness 1 30 Caucasian Master F English R 2 Asian Foreign 3 20 Bachelor 4 27 M 5 28 6 35 7 60 8 67 Beyond H.S L 9 PHD 10 70
Summary of handwriting samples 10 subjects Each subject wrote 3 authentic handwriting samples 3 forgeries of each of the other 9 subjects Total 300 handwriting samples 30 authentic 270 forgeries Total 900 database records One online and two resolutions offline for each handwriting sample
Speed Hypothesis Test H0(null hypothesis): the mean speed for the authentic and forged handwritings are about equal Ha (alternate hypothesis): the mean speed of the authentic handwriting is greater than that of the forged
Mean equality test output Alpha (level of significance) = 5% Authentic Forged Mean 0.083 0.057 Variance 0.00050 0.00053 Observations na=30 nf=270 Pooled Variance Hypothesized Mean Difference df 298 t Stat 5.87 P(T<=t) one-tail 5.90E-09 t Critical one-tail 1.65
Reject the null hypothesis Alpha (level of significance) = 0.05 p (probability) value is 5.90E-09 which is much less than alpha Reject null hypothesis with a 95% confidence interval Successfully prove the hypothesis
Wrinkliness Hypothesis Test H0 (null hypothesis): log2 ( 600dpif / 300dpif) ~ log2 ( 600dpia/ 300dpia) Ha (alternative hypothesis): the mean wrinkliness of the authentic handwriting is less than the mean wrinkliness of the forged handwriting
Mean equality test output Alpha (level of significance) = 5% Forged Authentic Mean 1.094 1.083 Variance 0.0013 0.0010 Observations 270 30 Pooled Variance Hypothesized Mean Difference df 298 t Stat 1.52 P(T<=t) one-tail 0.065 t Critical one-tail 1.65
Accept the null hypothesis Alpha (level of significance) = 0.05 p (probability) value is 0.065 which is greater than alpha Accept null hypothesis with 95% confidence interval Fail to prove the hypothesis
The first possible reason for failure Different writing styles among the three tries of the authentic handwriting First try Second try Third try
The second possible reason for failure Some subjects didn’t forge other subjects’ handwritings carefully Authentic Forged
Revised hypothesis test Eliminate the different authentic writing styles and the poorly forged handwriting samples Run the hypothesis test again
Mean equality test output Alpha (level of significance) = 5% Forged Authentic Mean 1.097 1.079 Variance 0.0016 0.0009 Observations 190 23 Pooled Variance 0.0015 Hypothesized Mean Difference df 211 t Stat 2.06 P(T<=t) one-tail 0.0205 t Critical one-tail 1.65
Reject the null hypothesis Alpha (level of significance) = 0.05 p (probability) value is 0.0205 which is less than alpha Reject null hypothesis with 95% confidence interval Successfully prove the hypothesis
Conclusion The average writing speed of the forged handwritings tends to be slower than the speed of the authentic handwritings “Good” (well formed) forged handwritings tend to be wrinklier (less smooth) than authentic ones
Future Extensions Redo the study using signatures rather than arbitrary words since writing signatures is a highly learned automatic process Investigate using different resolutions to improve the estimate of wrinkliness Devise pattern recognition algorithms to filter out the “bad” forged samples automatically Compute features over portions of the writing rather than over the whole word or signature
The End