GAFSV-Net: A Vision Framework for Online Signature Verification

1. Problem Statement

Online Signature Verification (OSV) aims to distinguish genuine signatures from skilled forgeries. This task faces two primary challenges:

1. High Intra-class Variability: A user's signature varies significantly across different writing instances due to speed, pressure, and environmental factors.
2. Data Scarcity: Verification systems often have very few enrollment samples (e.g., 1–4 signatures) per user, making it difficult to learn robust writer-independent models.

Limitations of Existing Methods:

• Classical Approaches: Rely on handcrafted features (e.g., DTW, GMM) which generalize poorly across writing styles.
• Deep Learning (1D): Most current deep learning models (RNNs, LSTMs, 1D-CNNs, Transformers) process raw temporal sequences directly. This restricts them to 1D architectures, preventing the use of powerful pretrained 2D vision backbones (e.g., those trained on ImageNet) and failing to leverage 2D spatial inductive biases.
• Gap: There is a lack of methods that effectively convert 1D time-series signature data into a 2D representation suitable for advanced computer vision models.

2. Methodology: GAFSV-Net

The authors propose GAFSV-Net, a dual-branch deep learning framework that bridges the gap between 1D temporal data and 2D vision architectures.

A. Data Representation: Asymmetric Gramian Angular Fields (GAF)

Instead of feeding raw $(x, y, p)$ sequences into a network, the method first transforms the data into 2D images:

1. Kinematic Extraction: Three time-series are derived from the raw stylus data:
   • Pen Speed ($v$)
   • Pressure Derivative ($\dot{p}$)
   • Direction Angle ($\theta$)
2. Encoding: Each of the three channels is independently encoded into two complementary 2D matrices:
   • GASF (Gramian Angular Summation Field): Encodes value co-occurrence (symmetric structure). It captures when two time steps have similar values.
   • GADF (Gramian Angular Difference Field): Encodes directional transitions (anti-symmetric structure). It captures whether a value is increasing or decreasing between time steps.
3. Asymmetric Construction: To preserve causal temporal ordering and reduce redundancy, the authors split the time series into two halves. The upper triangle of the matrix represents the first half's correlation with itself, and the lower triangle represents the second half. This results in a six-channel image ($3 \text{ channels} \times 2 \text{ variants}$) of size $M/2 \times M/2$.

B. Network Architecture

The framework utilizes a Dual-Branch ConvNeXt-Tiny architecture:

• Dual Backbones: Two independent ConvNeXt-Tiny networks (pretrained on ImageNet-1k) process the GASF and GADF halves separately. This allows each branch to learn filters specialized for the specific statistical properties of symmetric (GASF) vs. anti-symmetric (GADF) images.
• Attention Mechanisms:
  • Intra-branch Self-Attention: Refines feature maps within each branch to capture global spatial correlations.
  • Bidirectional Cross-Attention: Fuses information between the GASF and GADF branches. This allows the GASF branch to query discriminative patterns in the GADF branch (and vice versa) before projection.
• Projection: The fused features are projected onto a unit hypersphere to create a compact $D$-dimensional embedding.

C. Training Strategy

The model is trained using a writer-independent metric learning approach:

• Loss Function: A combination of Semi-hard Triplet Loss and a Uniformity Regularizer.
• Hard-Negative Injection: Skilled forgeries are explicitly injected as "hard negatives" during training. A specific cluster-level term ensures that forgeries are pushed away from their target writer's genuine cluster, not just from random writers.
• Uniformity: Prevents "embedding collapse" (where all samples cluster in one spot) by encouraging embeddings to spread evenly across the hypersphere.

D. Verification

At test time, the system computes the cosine similarity between a query signature embedding and the mean of the user's enrollment prototypes. If the score exceeds a threshold, the signature is accepted.

3. Key Contributions

1. Novel Representation: Proposes the first application of Asymmetric GASF/GADF encoding for OSV, transforming 1D kinematic signals into 2D images that explicitly map pairwise temporal correlations.
2. Dual-Branch Architecture: Demonstrates that GASF and GADF capture orthogonal aspects of temporal dynamics (co-occurrence vs. direction). The dual-branch design with cross-attention effectively leverages these complementary views.
3. Transfer Learning: Successfully transfers ImageNet-pretrained 2D vision backbones to the data-scarce OSV domain, outperforming models trained from scratch on 1D sequences.
4. Controlled Comparison: Provides the first fair comparison between GAF-based methods and sequence-based baselines (BiLSTM, Transformer, TCN) using identical training objectives, isolating the gain from the representation itself.

4. Experimental Results

The framework was evaluated on DeepSignDB and BiosecurID datasets.

• Performance:
  • BiosecurID: Achieved a 2.97% Equal Error Rate (EER) with 4 enrollment samples, significantly outperforming the previous best (DeepSign at 5.61%) and the standard DTW baseline (10.43%).
  • DeepSignDB: Achieved 6.01% EER, outperforming all sequence-based baselines (BiGRU, LSTM, Transformer, TCN) trained with the same objective.
• Comparison with Baselines:
  • GAFSV-Net consistently outperformed 1D models (RNNs, 1D-CNNs, Transformers).
  • It even outperformed MOMENT, a large time-series foundation model, proving that the 2D GAF encoding provides inductive biases that sequence-domain pretraining cannot recover.
• Ablation Studies:
  • GADF vs. GASF: GADF (directional) was found slightly more discriminative than GASF (co-occurrence) for skilled forgeries, but the combination yielded the best results.
  • Cross-Attention: The bidirectional cross-attention mechanism provided a ~0.6% improvement over simple concatenation, proving the value of dynamic feature fusion.
  • Loss Components: The combination of Triplet Loss, Forgery Injection, and Uniformity Regularization was essential; removing any component degraded performance significantly.

5. Significance

• Paradigm Shift: GAFSV-Net establishes that converting time-series data into 2D structured images allows OSV systems to leverage the massive representational power of pretrained computer vision models, a capability previously unavailable to 1D sequence models.
• Robustness: The method is particularly effective against skilled forgeries, which often mimic the gross shape of a signature but fail to replicate the fine-grained kinematic dynamics (acceleration, pressure changes) captured by the GADF encoding.
• Generalizability: The framework is writer-independent and requires no user-specific parameters at test time, making it suitable for real-world deployment with limited enrollment data.

In conclusion, GAFSV-Net demonstrates that 2D temporal encoding is a superior representation for online signature verification, offering a new direction for biometric security that moves beyond traditional 1D sequence modeling.