SIGN-GAIL: Rewarding Online Signature Generation for Digital Imitation

Anurag Pandey, Arnav Bhavsar Vinayak, Aditya Nigam, Divya Acharya, Balaji Rao K, Basu Verma; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1382-1391

In real-world scenarios signature generation involves mimicking a user's unique behavioral trajectory during online signing. This process presents a significant challenge: learning from expert data without direct interaction or explicit feedback. Inverse reinforcement learning (IRL) approaches attempt to address this by inferring an underlying reward function from expert data and using reinforcement learning (RL) to derive policies. However these methods are often slow and computationally expensive. To overcome these limitations we propose SIGN-GAIL a novel framework that leverages generative adversarial imitation learning (GAIL) to directly learn policies from expert data. Unlike traditional RL algorithms with manually defined reward functions GAIL trains a reward function adversarially enabling it to act as a discriminator to distinguish between expert and generated trajectories. By framing the problem as a generative adversarial task SIGN-GAIL effectively imitates complex behavioral trajectories achieving high fidelity in online signature generation. The proposed framework advances synthetic data generation in computer vision enhancing biometric authentication systems with robust dataset augmentation and improved resistance to deepfake forgeries. Experimental results show that SIGN-GAIL outperforms traditional methods in trajectory fidelity and resemblance demonstrating its potential for learning expert behaviors in sequential tasks like online signature generation.