Project about privacy in Machine Learning for the course 02460 Advanced Machine Learning Spring 2021 at Technical University of Denmark.
We explore the robustness and techniques of deep leakage from gradients (DLG) [1]. We investigate the influence of different initalization methods and distance measures by comparing the convergence rate and speed of single image reconstructions. Specifically, we compare the results with SAPAG [2].
[1] Zhu et. al (2020). "Deep Leakage from Gradients” Lecture Notes in Computer Science (including sub-series Lecture Notes in Artificial Intelligence and LectureNotes in Bioinformatics), vol. 12500 LNCS, no. NeurIPS,pp. 17–31.
[2] Wang et. al. (2020). “SAPAG: A self-adaptive privacy attackfrom gradients,”. arXiv.