Cryptography and Security (cs.CR)

Abstract: Artificial Intelligence (AI) has found wide application, but also poses risks due to unintentional or malicious tampering during deployment. Regular checks are therefore necessary to detect and prevent such risks. Fragile watermarking is a technique used to identify tampering in AI models. However, previous methods have faced challenges including risks of omission, additional information transmission, and inability to locate tampering precisely. In this paper, we propose a method for detecting tampered parameters and bits, which can be used to detect, locate, and restore parameters that have been tampered with. We also propose an adaptive embedding method that maximizes information capacity while maintaining model accuracy. Our approach was tested on multiple neural networks subjected to attacks that modified weight parameters, and our results demonstrate that our method achieved great recovery performance when the modification rate was below 20%. Furthermore, for models where watermarking significantly affected accuracy, we utilized an adaptive bit technique to recover more than 15% of the accuracy loss of the model.

Abstract: Physical persons playing the role of car drivers consume data that is sourced from the Internet and, at the same time, themselves act as sources of relevant data. It follows that citizens' privacy is potentially at risk while they drive, hence the need to model privacy threats in this application domain. This paper addresses the privacy threats by updating a recent threat-modelling methodology and by tailoring it specifically to the soft privacy target property, which ensures citizens' full control on their personal data. The methodology now features the sources of documentation as an explicit variable that is to be considered. It is demonstrated by including a new version of the de-facto standard LINDDUN methodology as well as an additional source by ENISA which is found to be relevant to soft privacy. The main findings are a set of 23 domain-independent threats, 43 domain-specific assets and 525 domain-dependent threats for the target property in the automotive domain. While these exceed their previous versions, their main value is to offer self-evident support to at least two arguments. One is that LINDDUN has evolved much the way our original methodology already advocated because a few of our previously suggested extensions are no longer outstanding. The other one is that ENISA's treatment of privacy aboard smart cars should be extended considerably because our 525 threats fall in the same scope.