Tian Dong (Shanghai Jiao Tong University), Shaofeng Li (Shanghai Jiao Tong University), Guoxing Chen (Shanghai Jiao Tong University), Minhui Xue (CSIRO's Data61), Haojin Zhu (Shanghai Jiao Tong University), Zhen Liu (Shanghai Jiao Tong University)

Identity plays an important role in responsible artificial intelligence (AI): it acts as a unique marker for deep learning (DL) models and can be used to trace those accountable for irresponsible use of models. Consequently, effective DL identity audit is fundamental for building responsible AI. Besides models, training datasets determine what features a model can learn, and thus should be paid equal attention in identity audit. In this work, we propose the first practical scheme, named RAI2, for responsible identity audit for both datasets and models. We develop our dataset and model similarity estimation methods that can work with black-box access to suspect models. The proposed methods can quantitatively determine the identity of datasets and models by estimating the similarity between the owner's and suspect's. Finally, we realize our responsible audit scheme based on the commitment scheme, enabling the owner to register datasets and models to a trusted third party (TTP) which is in charge of dataset and model regulation and forensics of copyright infringement. Extensive evaluation on 14 model architectures and 6 visual and textual datasets shows that our scheme can accurately identify the dataset and model with the proposed similarity estimation methods. We hope that our audit methodology will not only fill the gap in achieving identity arbitration but also ride on the wave of AI governance in this chaotic world.

View More Papers

Folk Models of Misinformation on Social Media

Filipo Sharevski (DePaul University), Amy Devine (DePaul University), Emma Pieroni (DePaul University), Peter Jachim (DePaul University)

Read More

Smarter Contracts: Detecting Vulnerabilities in Smart Contracts with Deep...

Christoph Sendner (University of Wuerzburg), Huili Chen (University of California San Diego), Hossein Fereidooni (Technische Universität Darmstadt), Lukas Petzi (University of Wuerzburg), Jan König (University of Wuerzburg), Jasper Stang (University of Wuerzburg), Alexandra Dmitrienko (University of Wuerzburg), Ahmad-Reza Sadeghi (Technical University of Darmstadt), Farinaz Koushanfar (University of California San Diego)

Read More

Securing Federated Sensitive Topic Classification against Poisoning Attacks

Tianyue Chu (IMDEA Networks Institute), Alvaro Garcia-Recuero (IMDEA Networks Institute), Costas Iordanou (Cyprus University of Technology), Georgios Smaragdakis (TU Delft), Nikolaos Laoutaris (IMDEA Networks Institute)

Read More

dewolf: Improving Decompilation by leveraging User Surveys

Steffen Enders, Eva-Maria C. Behner, Niklas Bergmann, Mariia Rybalka, Elmar Padilla (Fraunhofer FKIE, Germany), Er Xue Hui, Henry Low, Nicholas Sim (DSO National Laboratories, Singapore)

Read More