gradient-inversion

[NDSS 2025] CENSOR: Defense Against Gradient Inversion via Orthogonal Subspace Bayesian Sampling

[ICCV 2025] Geminio is a VLM-powered gradient inversion attack in federated learning (FL). It allows the adversary (the FL server) to describe the data of value and reconstruct the victim client's private data matching the description.

Federated learning + iDLG gradient inversion attack + Central DP defense + Gradio demo. The honest finding: naive Central DP collapses utility on small federations (Gaussian-mechanism curse of dimensionality). Production fixes (DP-SGD, Opacus) documented.

Companion tooling for capturing and reconstructing gradients from federated ASR clients.

A comprehensive framework simulating integrity (Backdoor) and privacy (Gradient Inversion) attacks in Federated Learning, featuring a Machine Unlearning defense strategy

Exact recovery of speech from federated ASR updates: the first-layer gradient spans the client's feature windows, and the overlap makes them solvable in closed form. No transcript, no optimisation, up to 35 s.

Differential Privacy & Gradient Defense in Secure Federated Averaging — MSc Thesis, ELTE University

Superseded by takakhoo/unsplice. Earlier optimisation-based gradient inversion for CTC speech models.

An implementation of the paper "Cascading Gradient Inversion via LT-Code Inspired Peeling in Federated Learning"

From-scratch DLG/iDLG gradient inversion attacks and Bonawitz Secure Aggregation on federated MNIST. The attack works, the defence stops it at no accuracy cost.

Gradient inversion attacks and privacy defenses in federated learning using iDLG, PRECODE, and homomorphic encryption.