A. K. Mandal

Privacy-preserving methodologies against privacy attacks on deep learning: a survey

The analysis finds that noise-based methods such as DP remain the practical baseline but offer only partial protection; cryptographic approaches provide stronger theoretical guarantees at substantially higher cost; and LLM/multimodal leakage remains an urgent, under-benchmarked gap.