Detection of Information Hiding at Physical Layer in Wireless Communications

This article concerns the problem of detecting the use of information hiding at the physical layer in wireless communications because they are harder to be detected than other layer-based information hiding schemes. Prior schemes for detecting physical layer based information hiding are either heuristic based or machine learning based. The key limitation of prior heuristics based information hiding detection schemes is that they do not answer the fundamental question of why the information hidden at the physical layer can be detected. The key limitation of prior machine learning based information hiding detection schemes is that they lack robustness because wireless signals at the physical layer are very much environmental dependent, and thus an information hiding detection scheme trained in one environment often does not work well in another environment. Our insight is that embedding information on wireless signals at the physical layer will inevitably have a negative impact on the decodability of the cover signals, such as the increase of the error probability at the receiver (as well as the monitor). Based on the above insight, in our approach, after the monitor demodulates and decodes the cover signals, it will re-encode and re-modulate the cover signals, and then compare the resulting recovered signals with the raw signals that it received from the sender. We further propose a new estimation scheme for calculating receiver noise variance and conducted theoretical analysis. Specifically, we propose two hidden information detection schemes, a noise grouped based detection scheme and a constellation distance based detection scheme, both taking estimation errors into consideration. In particular, our constellation distance based detection scheme is the first scheme that can pinpoint the exact location on the received signals that are embedded with hidden information. We implemented our schemes and conducted extensive performance comparison between our schemes and prior schemes. Our experimental results show that when the received SNR is more than 20 dB, our approach with the new estimation scheme has the probability of detection more than 0.95, and our constellation distance based detection scheme can correctly pinpoint all embedded locations.