A0935
Title: Exploring the benefits of visual prompting in differential privacy
Authors: Xuebin Ren - Xi\'an Jiaotong University (China) [presenting]
Abstract: Visual prompting (VP) is an emerging and powerful technique that allows sample-efficient adaptation to downstream tasks by engineering a well-trained frozen source model. The benefits of VP are explored in constructing compelling neural network classifiers with differential privacy (DP). VP is explored and integrated into canonical DP training methods, and its simplicity and efficiency are demonstrated. In particular, it is discovered that VP, in tandem with PATE (i.e., a state-of-the-art DP training method that leverages the knowledge transfer from an ensemble of teachers), achieves the state-of-the-art privacy-utility tradeoff with minimum expenditure of privacy budget. Moreover, additional experiments on cross-domain image classification are conducted with a sufficient domain gap to further unveil the advantage of VP in DP. Lastly, extensive ablation studies are also conducted to validate the effectiveness and contribution of VP under DP consideration.
