Paper accepted at ACM CIKM 2026

29 May 2026

We are excited to share our new paper, “Revealing Hidden Vulnerabilities in Autoencoders through Gradient Signal Restoration,” accepted at the 35th ACM International Conference on Information and Knowledge Management (CIKM 2026).

This work is authored by Chethan Krishnamurthy Ramanaik, Arjun Roy, Tobias Callies, and Prof. Eirini Ntoutsi.

Autoencoders are widely used to learn compressed representations and reconstruct complex data. However, their robustness against adversarial perturbations remains less understood than that of conventional classification models.

The paper shows that existing attacks can underestimate autoencoder vulnerability because adversarial gradients may vanish as they pass through ill-conditioned intermediate layers. As a result, attacks can converge to weak perturbations and create an overly optimistic impression of model robustness.

Key takeaways:

  • Hidden vulnerabilities. Near-zero singular values in intermediate weight matrices can degrade adversarial gradient signals and weaken existing attacks.

  • Gradient restoration. The proposed GRILL framework restores these signals during optimization, allowing attacks to find more damaging perturbations under the same constraints.

  • More reliable robustness evaluation. GRILL significantly improves attack effectiveness across multiple autoencoder architectures.

  • Broad experimental validation. The method is evaluated for sample-specific and universal attacks, as well as under standard and adaptive attack settings.

  • Relevance beyond autoencoders. Initial results suggest that similar vulnerabilities may also occur in modern multimodal encoder-decoder architectures.

The work highlights the importance of reliable attack optimization when evaluating robustness: an apparently robust model may simply be protected by weaknesses in the evaluation method.