DimShield: Exploring Intrinsic Dimension Estimation for Enhanced Machine Learning Security

DimShield is a runtime defense framework that uses intrinsic dimension estimation to identify adversarial inputs to machine learning models deployed in security-critical settings. By analyzing the geometry of the feature space at inference time, DimShield can detect adversarially crafted inputs — including those designed to evade microarchitectural attack detectors — without requiring retraining or access to the attack methodology. This work complements existing adversarial defenses with a theoretically grounded, architecture-agnostic approach.

Recommended citation: D. R. Dipta, K. Christofferson, S. Seonghun and B. Gulmezoglu, "DimShield: Exploring Intrinsic Dimension Estimation for Enhanced Machine Learning Security." Under review.