DeepAuth: A DNN Authentication Framework by Model-Unique and Fragile Signature Embedding

• DOI: 10.1609/aaai.v36i9.21193
• Publication Date: 2022-07-04T09:52:09Z
• Authors: Yingjie Lao; Weijie Zhao; Peng Yang; Ping Li
• Keywords: Backdoor; Signature (topology)

Along with the evolution of deep neural networks (DNNs) in many real-world applications, complexity model building has also dramatically increased. Therefore, it is vital to protect intellectual property (IP) builder and ensure trustworthiness deployed models. Meanwhile, adversarial attacks on DNNs (e.g., backdoor poisoning attacks) that seek inject malicious behaviors have been investigated recently, demanding a means for verifying integrity users. This paper presents novel DNN authentication framework DeepAuth embeds unique fragile signature each protected model. Our approach exploits sensitive key samples are well crafted from input space latent then logit producing signatures. After embedding, will respond distinctively these samples, which creates model-unique as strong tool user identity. The embedding process designed fragility signature, can be used detect modifications such an illegitimate or altered should not intact signature. Extensive evaluations various models over wide range datasets demonstrate effectiveness efficiency proposed DeepAuth.