Model-Free Undetectable Attacks on Linear Systems Using LWE-Based Encryption

We show that the homomorphic property, a desired property in encrypted control, can lead to failure in the cyber defense of a dynamical control system from undetectable attacks, even though individual signal sequences remain unknown to the attacker. We consider an encryption method based on the Learning with Errors (LWE) problem and demonstrate how model-free undetectable attacks on linear systems over integers can be computed from sampled inputs and outputs that are encrypted. Previous work has shown that computing such attacks is possible on nonencrypted systems. Applying this earlier work to our scenario, with minor modifications, typically amplifies the error in encrypted messages unless a short vector problem is solved. Given that an attacker obtains a short vector, we derive the probability that the attack is detected and show how it explicitly depends on the encryption parameters. Finally, we simulate an attack obtained by our method on an encrypted linear system over integers and conduct an analysis of the probability that the attack will be detected.
QC 20230307