1Professor, HOD, Department of CSE, KGiSL Institute of Technology, Coimbatore, 641035, India. Email: hodcse@kgkite.ac.in
2Department of CSE, KGiSL Institute of Technology, Coimbatore, 641035, India. Email: mvkkumar2006@gmail.com
3Department of CSE, KGiSL Institute of Technology, Coimbatore, 641035, India. Email: divibalakrishnan2005@gmail.com
4Department of CSE, KGiSL Institute of Technology, Coimbatore, 641035, India. Email: hari20041203@gmail.com
5Department of CSE, KGiSL Institute of Technology, Coimbatore, 641035, India. Email: indreshx2@gmail.com
Face recognition systems have been propelled to new heights by the deep learning. You can now see them everywhere in regulating access and policing crowds, monitoring attendance. The majority of studies narrow down to making such systems more precise and solid. However, frankly, the security and integrity of the training data upon which they work is not discussed sufficiently by people. Deep learning systems are heavily reliant on their data, and, therefore, in case an individual alters or poisons the data, the entire machine can silently crash. Credibility flies down the drain. DeepAudit comes at that point. It is a framework that ensures that face recognition data is not tampered and poisoned. As opposed to tampering with the model or altering its learning process, DeepAudit places tiny, imperceptible watermarks in the facial embeddings during enrollment. These cues do not disrupt performance of recognition in any way. However, when you have to verify the integrity of your data or audit your dataset in the future, those watermarks are available. Here is the mechanism: First during enrollment, a neural net converts facial images into numberical representations. DeepAudit watermarks each of them and stores them safely. Then when recognition is required, the system simply compares live faces to such stored embeddings no muss no messing with the watermark. The watermark will not interfere until it is necessary to check data integrity or audit it. The tests indicate that these embedded watermarks do not affect recognition accuracy or similarity scores. Nevertheless, DeepAudit can detect tampering of the data by an individual. The main takeaway? To be serious about biometric security, it is essential that you are concerned about the integrity of data, and not plain accuracy. DeepAudit provides you with a scalable and practical means of ensuring that face recognition systems are trustworthy on the inside.
Keywords: Adversarial Robustness, Identity Verification Systems, Biometric Template Security, Watermark-Based Protection, Integrity-Aware Recognition, Deep Facial Embeddings, and Data Poisoning Mitigation
How to cite this article: Thenmozhi T, Abhinaya V, Divya B, Harikrishnan P, Indresan V., DeepAudit: An Integrity-Aware Deep Facial Recognition System Using Watermarked Embedding...Int J Drug Deliv Technol. 2026; 16(11s): 726-744; DOI: 10.25258/ijddt.16.11s.75
Source of support: Nil.
Conflict of interest: None