Explore To Mimic: A Reinforcement Learning Based Agent To Generate Online Signatures

Anurag Pandey, PushapDeep Singh, Arnav Bhavsar, Aditya Nigam, Dr. Divya Acharya, Basu Verma

Keywords: Reinforcement Learning, Online Signature, Biometric, Generative Model, On-Policy

Abstract: Recent advancements in utilising decision making capability of Reinforcement Learning (RL) have paved the way for innovative approaches in data generation. This research explores the application of model free on-policy RL algorithms for generating online signatures and its controlled variations. Online signatures are captured via e-pads as sequential structural coordinates. In this study, we have introduced a robust on-policy RL agent named as SIGN-Agent, capable of generating online signatures accurately. Unlike other RL algorithms, on-policy RL directly learns from the agent's current policy, offering significant advantages in stability and faster convergence for sequential decision-making. The proposed SIGN-Agent operates in a random continuous action space with controlled exploration limits, allowing it to capture complex signature patterns while minimizing errors over time. The downstream applications of this system can be extended in diverse fields such as enhancing the robustness of signature authentication systems, supporting robotics, and even diagnosing neurological disorders. By generating reliable, human-like online signatures, our approach strengthens signature authentication systems by reducing susceptibility towards system-generated forgeries, if trained against them. Additionally, the proposed work is optimized for low-footprint edge devices, enabling it to function efficiently in the area of robotics for online signature generation tasks. Experimental results, tested on large, publicly available datasets, demonstrate the effectiveness of model free on-policy RL algorithms in generating online signature trajectories, that closely resemble user's reference signatures. Our approach highlights the potential of model free on-policy RL as an advancement in the field of data generation targeting the domain of online signatures in this research.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11448