Hyperspectral imaging is a popular style of image acquisition which captures a multitude of two-dimensional images at different frequencies to yield more information from a scene. For imaging in certain frequency ranges, point-scanning images pixel-by-pixel is the only option for imaging and consequently can require prohibitively long measurement times. This paper introduces a non-uniform compressed sensing strategy for recovering whole hyperspectral images with much fewer samples than normally required. Our methods use the known variability of measurements across training subjects to weight random sampling toward unreliable hyperspectral points. This strategy is used to heavily reduce remote hyperspectral image acquisition time while maintaining or improving recovery performance. To test our methods, we focus on the problem of recovering hyperspectral images for non-destructively detecting malicious dormant hardware Trojans hidden in integrated circuits. These non-destructive detection techniques require remote hyperspectral measurements of the backscattering electromagnetic side channel; which can only be performed by point-scanning with existing technologies. We detect covert hardware Trojan circuit modifications with state-of-the-art performance while requiring up to ten times fewer measurements than prior methods. We compare performance of a uniform-random and our weighted-random compressed sensing strategy for the 2-dimensional discrete cosine transform bases as well as with learned dictionary bases.
INDEX TERMSCompressed sensing, non-uniform sampling, hyperspectral imaging, hardware Trojan, hardware security, backscattering EM side channel. I. INTRODUCTION 19 In the past two decades, compressed sensing (CS) has evolved 20 from a novel theoretical outline to become a popular frame-21 work to recover high-dimensional images from relatively few 22 samples in a variety of problem domains. Under the assump-23 tion that the data of interest can be represented sparsely in 24 some basis, CS enables the accurate recovery of signals even 25 when sampling far less than Shannon-Nyquist theory would 26 imaging in a variety of fields such as archaeology [11], 49 geoscience [12], biomedicine [13], [14], and hardware secu-50 rity [15], [16]. HSI involves acquiring potentially hundreds 51 of two-dimensional images across a range of frequencies 52 and can require long capture times or complicated hardware. 53 Since the development of the compressed sensing frame-54 work, hyperspectral imaging (HSI) has received significant 55 attention as a target for CS due to time required to scan 56 the high-dimensional feature space in which a hyperspectral 57 image exists. Early works used coded apertures to capture 58 transformed images through random pixel masks paired with 59 assumptions of spatial smoothness to reconstruct hyperspec-60 tral images [17], [18]. These physically project incoming 61 light from a scene to a compressed space directly, which is 62 then reconstructed algorithmically. Instead of hyperspectral 63 imaging wit...