SIGuard Guards Secure Inference Against MIAs
Researchers from RMIT University, CSIRO Data61 and University of Melbourne present SIGuard at NDSS 2025, a framework defending MPC-based secure inference against membership inference attacks. SIGuard perturbs encrypted model predictions via an MPC and machine-learning co-design, reducing attack accuracy to near random while adding 1.1s overhead and occupying about 24.8% of a 3.29s ResNet34 CIFAR-10 inference. The protocol integrates with existing MPC pipelines without degrading accuracy.
Key Points
- 1Demonstrates MPC-based secure inference remains vulnerable to membership inference, with attack success comparable to plaintext
- 2Introduces SIGuard that perturbs encrypted predictions via MPC-machine-learning co-design to protect output privacy
- 3Reduces MIA accuracy to near-random while adding 1.1s overhead, applicable to ResNet34 CIFAR-10 evaluations
Scoring Rationale
High novelty and practical defense with strong NDSS credibility; limited scope focused on MPC-based secure inference.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
