Supervised Machine Learning with Plausible Deniability. Issue 112 (January 2022)

• Record Type: Journal Article
• Authors: Rass, Stefan; König, Sandra; Wachter, Jasmin; Egger, Manuel; Hobisch, Manuel
• Is Part Of: Computers & security. Issue 112(2022)
• Publication Date: 2022-01
• Subjects: Machine Learning -- Plausible Deniability -- Privacy -- Data Protection -- Artificial Intelligence; Computer security -- Periodicals; Electronic data processing departments -- Security measures -- Periodicals; 005.805
• DOI: 10.1016/j.cose.2021.102506
• Languages: English
• ISSNs: 0167-4048

Abstract

Highlights: Introduction of the new concept of "deniability" and "plausible deniability" in the context of machine learning, as a notion of security in this context Formally proven conditions under which plausible deniability applies (Theorems 1 and 2) Numerical evaluation of the concepts, showcasing it on several machine learning models Discussion on an extended level about other machine learning models to which our results do not apply to (to show the scope of our work) Derivation of implications and meaning for de facto standard methods of training machine learning, culminating in recommendations on how to practically train models to avoid security concerns connected to plausible deniability Graphical abstract: Abstract: We study the question of how well machine learning (ML) models trained on a certain data set provide privacy for the training data or, equivalently, whether it is possible to reverse-engineer the training data from a given ML model. While this is easy to answer negatively in the most general case, it is interesting to note that the protection extends beyond non-recoverability towards plausible deniability : Given a ML model f, we show that one can take a set of purely random training data, and from this define a suitable "learning rule" that will produce a ML model that is exactly f . Thus, any speculation about which data has been used to train f is deniable upon the claim that any other data could have led to the same results. We corroborate our theoretical finding with practical examples and open source implementations of how to find the learning rules for a chosen set of training data.