Robust Binary Quantizers for Distributed Detection

We consider robust signal processing techniques for inference-centric distributed sensor networks operating in the presence of possible sensor and/or communication failures. Motivated by the multiple description (MD) principle, we develop robust distributed quantization schemes for a decentralized detection system. Specifically, focusing on a two-sensor system, our design criterion mirrors that of MD principle: if one of the two transmissions fails, we can guarantee an acceptable performance, while enhanced performance can be achieved if both transmissions are successful. Different from the conventional MD problem is the distributed nature of the problem as well as the use of error probability as the performance measure. Two different optimization criteria are used in the distributed quantizer design, the first a constrained optimization problem, and the second using an erasure channel model. We demonstrate that these two formulations are intrinsically related to each other. Further, using a person-by-person optimization approach, we propose an iterative algorithm to find the optimal local quantization thresholds. A design example is provided to illustrate the validity of the iterative algorithm and the improved robustness compared to the classical distributed detection approach that disregards the possible transmission losses.