Multi-class Traffic Morphing for Encrypted VoIP Communication

In a re-identification attack, an adversary analyzes the sizes of intercepted encrypted VoIP packets to infer characteristics of the underlying audio— for example, the language or individual phrases spoken on the encrypted VoIP call. Traffic morphing has been proposed as a general solution for defending against such attacks. In traffic morphing, the sender pads ciphertext to obfuscate the distribution of packet sizes, impairing the adversary’s ability to accurately identify features of the plaintext. This paper makes several contributions to traffic morphing defenses. First, we argue that existing traffic morphing techniques are ineffective against certain reidentification attacks since they (i) require a priori knowledge of what information the adversary is trying to learn about the plaintext (e.g., language, the identity of the speaker, the speaker’s gender, etc.), and (ii) perform poorly with a large number of classes. Second, we introduce new algorithms for traffic morphing that are more generally applicable and do not depend on assumptions about the goals of the adversary. Finally, we evaluate our defenses against re-identification attacks, and show, using a large real-world corpus of spoken audio samples, that our techniques reduce the adversary’s accuracy by 94% with low computational and bandwidth overhead.