Transcript Slide 1
CSE - 717 Introduction to Online Signature Verification Swapnil Khedekar Signature Verification Biometric Technology that verifies a user's identity by measuring a unique-to-the-individual biological trait Creates trust by establishing a context of confident privacy and undeniable personal responsibility Future and destiny of computerized network security and identification is Biometrics Signature verification Behavioral biometrics Verify user signatures using computers or embedded devices Efficient and effective method of replacing insecure passwords, PIN numbers, keycards and ID cards Why Signatures? Advantages Customary way of identity verification Even advanced PDAs focus pen-input People are willing to accept a signature based verification Easier, faster, low FRR, low memory Disadvantages Dynamic Biometric, Non-repudiation Can be forged easily Individuality Physiology studies suggest Handwriting originates & develops in brain Signal to duplicate mental picture of character or word is sent to the arm and hand Handwriting system = Machine General or class characteristics Shoulder, arm, hand, fingers work as levers and fulcrums During learning, signals are sent back to brain Strength & flexability of muscles, position of pen-grip and the overall posture of the writer all affect the output Mental state, writing instrument, surface etc also affect Thus, each person has a small range of natural variation General: Effect of culture, trend, teacher’s style etc Class: Conscious/unconscious individual changes Axiom A person is unlikely to ever duplicate any signature exactly Difference Static/Offline Early 1970’s Only image of signature No need of special hardware, ubiquitous use Large storage Can not trace speed, style, pressure etc Easier to forge Around 95% accuracy Dynamic/Online Early 1990’s Uses shape, speed, pressure Needs special digital surface, pads and pen etc. Numeric data, small storage Can use speed, pressure, angle of pen etc to further exploit individuality Harder to forge Around 99% accuracy •[Rigoll98] performed systematic comparison of online-offline techniques & their performance. Concluded with preference for on-line verification system. Capture Devices Technology SignatureGem Pressure sensitive sensors arranged in compact grid to form flat surface When pen touches a sensor, pressure at that sensor is calculated The sensors are scanned periodically for pen positions Position of sensor, pressure, pen angle are stored Periodic scanning results in sequence of parameters SigLite ClipGem ePad-ID Issues People use full names, initials or complex signs People tend to vaguely write ending part, dots etc Signatures on bank cheques & delivery books [Herbst99] showed trained experts can have 0% FAR, 25% FRR. Untrained have upto 50% FAR. [Osborn29] claimed many characteristics of natural writing can never be forged Also suggested that samples should be collected over time, not at single time [Hilton92] claimed single-most important feature is movement Typical System Reference signature: Matching Data acquisition Pre-processing Feature extraction A distance metric criteria is assumed Distance between test and reference signature is calculated If distance < threshold, it is authenticated Performance Evaluation On skilled and random forgeries No public standard signature dataset Features Used Features for online signatures Total time Signature path length Path tangent angles Signature velocity Signature accelerations Pen-up times & durations [Crane83] proposed 44 while [Parks85] proposed 90 features [Lee96] used 15 static & 34 dynamic None related to shape 1% FRR, 20% FAR on timed forgeries Distance Functions Linear Discriminant function Euclidian Distance Classifier Linear combination of features fi G(x) = wtx + w0, w=weighing vector,w0=class const Some researchers proposed feature vector normalized by reference mean ri or std. deviation si G(T) = (1/n) ∑ ( (ti – ri) / si )2 Least distant value is compared with threshold Synthetic Discriminant Matching Mostly used as post-processor in combination Finds filter impluse response w from samples Proposed by [Wilkinson90] and [Bahri88] Distance Functions Dynamic Programming Matching Minimize the residual error between two functions by finding a warping function Rescales one of original functions time axis Majority Classifier Main drawback of previous techniques FAR -> 100% as FRR -> 0% & vice versa Single distant feature influences other close features Genuine if atleast half features pass test Hidden Markov Models [Kashi98] Creates a universal prototype for signature, new signature is assigned a distance from the prototype Uses 21 Global & 5 local features Segmentation, parameter re-estimation done by the Viterbi 1% FRR, 2.5% FAR Distance Functions Multi-expert System [DiLeece00] 3 independent agents. Result by majority 3.2% FRR, 0.55% FAR with 3.2% undecided Velocity-based Models [Nalwa97] Shape-based features and holistic analysis Speed-based features Regional Analysis Velocities are hard to copy, good forgery detectors Look at both local and global models Weighted and biased harmonic mean as a way of combining errors from multiple models 2-5% error rate Split-and-Merge [Lee97] Static and dynamic features, Polar coordinates For Chinese signatures Splits into 2 parts & evaluate each & then combines results 13% FAR, 3% FRR Distance Functions Deformable structures [Pawlidis98] Neural networks [Paulik99] Signature identification instead of signature verification Focus on an active vision system Only orientation normalization, no size Attempt to create a vague outline to classify easily 2.8% false recognition. But 18.3% inconclusive Illustrates the difference in error by skilled versus random forgeries Random : 0.25% FAR & FRR. Skilled:2.3% FAR & 7% FRR. Curve aligning [Sebastian03] Compares the curves using an alignment curve Edit distance on length and curvature for aligning Alignment curve created a from prototype of each segment Software products PenOp Sign-On For verifying static signatures on cheques Cadix ID-007 For online signature login Dynamically updates reference signatures 2.5% FRR & FAR Signer confidence Peripheral Vision Use can login only using handwritten signatures Online signature verification in less than 1 sec CounterMatch Claims to match signature in any language Software products Kappa ApproveIT Signature added to WordPerfect document directly from penbased input If content of document are changed, signature won’t appear Unipen Uses “user-specific” features for lower FRR Tested on 8500 postal images. 0.85% FRR Look for regularities and lawfulness in writing Groups strokes together on a self-associating graph Looks at predecessor and successor strokes More similar to Handwriting Recognition Others SignCrypt, Q-Lock, Cyber-Sign Data transfer Storage & Retrieval [Han97] For Signature identification, can be extended for verification Codes features of the signature into a string Enters into database based on a hash-code of string Loops end, branch, convex, concave points used Proposed fast and efficient way of comparing and indexing these strings Conclusion The new system should be an on-line system Shape is an integral part of signature verification, it is a metric that is most easily imitated by a forger Both global & local features should be used Different methods have been tried with varying results, About 99% at the best Great deal of speed improvement to be done Signature segmentation into individual strokes needs attention Multi-expert system to integrate different methods Analysis on proper setting of thresholds & use of user-specific thresholds Sensors have developed to a fair point of saturation Study on multi-lingual signatures is unfocused