Generative Lossy Sensor Data Reconstructions for Robust Deep Inference

Limited communication bandwidth in modern net-works of wireless sensors, popularly known as the Internet of Things, may require lossy compression of the data. The sensor data, intended for inference by a deep learning (DL) model, will be reconstructed with some distortion by the remote user who received its compressed representation over a wireless channel. We study the robustness of the remote DL (RDL) model to the distortion-inducing lossy compression of the input, including the robustness under an adersarial attack. We are particularly interested in a novel data compression, known as learned compression (LC) due to the use of DL. Starting from MNIST images, we compare conventional compression (JPEG) with a published generative LC model under different compression ratios (CR). The generative LC, a hierarchical vector-quantized variational autoencoder, is state-of-the-art. With a CR up to 4 times that of JPEG’s, its reconstructions are achieving the same accuracy on a RDL MNIST classifier as with JPEG or uncompressed data. Also, RDL accuracy degrades gracefully with LC-induced information loss up to a remarkable CR. High compression allows for multiple generative LC descriptions: a single image generates many conditionally independent compressed representations of the same low rate. Their decompression creates randomized image reconstructions contributing the salient features needed in downstream RDL, and making it more robust.