Offline Handwritten Signature Verification Using a Stream-Based Approach

TL;DR

This paper proposes a novel HSV approach with an adaptive system that receives an infinite sequence of signatures and is updated over time, and demonstrates the superior performance of the proposed method compared to standard approaches that use a Support Vector Machine as a classifier.

Abstract

Handwritten Signature Verification (HSV) systems distinguish between genuine and forged signatures. Traditional HSV development involves a static batch configuration, constraining the system's ability to model signatures to the limited data available. Signatures exhibit high intra-class variability and are sensitive to various factors, including time and external influences, imparting them a dynamic nature. This paper investigates the signature learning process within a data stream context. We propose a novel HSV approach with an adaptive system that receives an infinite sequence of signatures and is updated over time. Experiments were carried out on GPDS Synthetic, CEDAR, and MCYT datasets. Results demonstrate the superior performance of the proposed method compared to standard approaches that use a Support Vector Machine as a classifier. Implementation of the method is available at https://github.com/kdMoura/stream_hsv.

Figures (6)

• Figure 1: Stream HSV system. $(S_C,\hat{l}_C)$ denotes a claimed signature from the stream, and $(S_R,l_R)$ a reference signature of the corresponding user. Signatures, after being preprocessed, have their features extracted by a representation model $\phi$. The stream dichotomy transformation is applied to the pair of features vectors $DT(\textbf{x}_R, \textbf{x}_C)$ and passed to the adaptive classifier $\theta$, which outputs a prediction. At the end, a fusion function is employed considering all reference signatures to deliver a final result. If true labels are available, the classifier is updated with all new dissimilarities information.
• Figure 2: Data segmentation into development $\mathcal{D}$ and exploitation $\mathcal{E}$ sets. To generate $\mathcal{E}$, a set of references $\mathbb{S}_R$ and claimed signatures $\mathbb{S}_C$ are randomly selected for all $n\mathcal{E}$ users. $\mathbb{S}_C$ contains genuine, random forgery, and skilled forgery samples. To generate $\mathcal{D}$, a set of genuine $\mathbb{S}_{G_\mathcal{D}}$ and random forgery $\mathbb{S}_{RF_\mathcal{D}}$ are randomly chosen for all $n\mathcal{D}$ users. Selected samples are utilized to perform dissimilarity transformations as defined in Equations \ref{['eq:D_set_pos']}, \ref{['eq:D_set_negative']}, \ref{['eq:E_set_user']}, and \ref{['eq:E_set']}.
• Figure 3: Stream $\mathbb{T}$ of claimed signatures obtained from the exploitation set $\mathcal{E}$. The number of users $n\mathcal{E}$ is represented by $n$, while the number of claimed signatures $n\mathcal{E}_C$ is denoted by $m$. $\mathbb{S}^{i,j}_C$ is the $j$-th set of claimed signature of user $i$, where $\mathbb{S}^{i,j}_C = \{S^{i,j}_G,S^{i,j}_{RF},S^{i,j}_{SK}\}$, with $S^{i,j}_G \in \mathbb{S}^i_{G_\mathcal{E}}$, $S^{i,j}_{RF} \in \mathbb{S}^i_{RF_\mathcal{E}}$ and $S^{i,j}_{SK} \in \mathbb{S}^i_{SK}$. G: genuine, RF: random forgery, SK: skilled forgery signature.
• Figure 4: Skilled forgery detection in batch settings for different development sets. #U and #G denote, respectively, the number of users $n\mathcal{D}$ and genuine signatures $n\mathcal{D}_G$ employed during training. #S denotes the resulting number of samples.
• Figure 5: Stream evaluation of skilled forgery detection on GPDS Synthetic using SVM and the Stream HSV (SHSV). The evaluation considers 12 reference signatures with a Max fusion for decision-making. SHSV is updated after every training chunk, employing only genuine and random forgery signatures, and evaluated on every window (Table \ref{['tab:stream_evaluation']}). #Users ($n\mathcal{D}$) and #G denote the number of users and genuine signatures used in the initial training, respectively. The horizontal red line shows the result ($7.93$) reported in Talles2023_AmultitaskApproach4ContrastiveLearning for 12 reference signatures, #G=12, and $n\mathcal{D}$ = 2000.