XAI Attack & Defense
Robustness of explanations under adversarial manipulation
- 91%
- drift reduction
- 246K+
- samples
- 3
- security datasets
- 4
- defense architectures
“A perturbation can leave a prediction untouched while completely shifting which features the explanation highlights.”
Phishing, intrusion and fraud detectors are black boxes, and the XAI methods introduced to open them can themselves be attacked. An analyst who trusts a manipulated explanation reaches the wrong conclusion while the model looks correct.
A research framework that trains base models across three real security datasets, generates explanations via four XAI methods, executes four families of explanation attacks, and quantifies manipulation with a formal Explanation Drift metric. It then analyses amplified vulnerability under few-shot learning and proposes four progressively stronger defense architectures, culminating in a hybrid defense.
What makes it
work.
Four XAI methods — SHAP, LIME, Integrated Gradients and Captum attributions — under four attack families.
A formal Explanation Drift metric that makes manipulation measurable rather than anecdotal.
Few-shot analysis showing where low-data regimes amplify explanation vulnerability.
Hybrid defense reduces explanation drift by up to 91% while preserving prediction accuracy.
Built with.
- PyTorch
- SHAP
- LIME
- Integrated Gradients
- Captum
- NumPy
Collaborators on the original technical report: Gokul Ram K, Kishore A G.