Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{206823,
        author = {Dinesh Naik},
        title = {Topology-Guided Adversarial Detection and Robustification of Deep Neural Networks Using Persistent Graph Signatures},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {no},
        pages = {594-603},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=206823},
        abstract = {Although deep neural networks have shown great success on image classification and pattern recognition problems, the networks are extremely susceptible to adversarial examples created by a small and hard-to-detect input perturbation. Current detection approaches mostly focus on either statistical feature distributions, confidence scores, or intermediate representation of activation. However, these methods do not necessarily reflect structural changes within the network during test time. The proposed topology-guided adversarial detection and robustification framework is motivated by recent findings that clean and adversarial inputs to neural networks flow through distinct paths, and leverages persistent graph signatures. In the proposed method, the neural network is modelled as an activation graph, triggered by the input, whose edge weights are obtained from the interaction between neuron activations and learned parameters of the model. Magnitude based parameter variation is used to detect edges that are under optimized and then persistent homology is used to obtain signatures of topology from the activation subgraphs. These signatures are then utilized for adversarial detection through a light weight Classifier. Further, a topology-guided edge-pruning method is proposed, which aims at mitigating adversarial risk by pruning structurally unstable and weakly optimized paths. The proposed framework is tested with benchmark image datasets under common adversarial attacks like FGSM and PGD. The experimental analysis will be expected to show that the persistent graph signatures are discriminative and able to distinguish between clean and adversarial samples, and that topology-guided pruning boosts model adversarial robustness with minimal or insignificant effects on clean accuracy. The proposed work provides an interpretable and structurally-grounded method for adversarial detection and defence in deep neural networks.},
        keywords = {Adversarial robustness, Topological data analysis, Persistent homology, Adversarial detection, Edge pruning.},
        month = {July},
        }