Concerning the problem that the transmission and storage of massive redundant image data in Internet of Things （IoT） lead to high resource consumption and potential privacy leakage， a Deep Compressive Sensing （DCS） network for IoT images and its chaotic encryption protection method were proposed. Firstly， an improved DCS network was proposed to achieve high-quality image compression and reconstruction. In this network， residual blocks in conventional deep reconstruction network were modified using a channel attention mechanism， and a parallel fusion design integrating multi-scale branches and a fusion module was adopted， thereby improving the reconstruction performance of traditional deep reconstruction network based on residual stacking in convolutional layers. Secondly， a multi-cavity chaotic system was proposed to realize spherical cavity expansion in the X， Y， or Z direction — either in one direction or both directions — through spherical coordinate transformation and two sets of parity-controlled step functions. This system has chaotic properties and randomness， making it suitable for image encryption. Finally， based on the proposed DCS network and the multi-cavity chaotic system， encryption and decryption for DCS measured images was designed using chaotic index scrambling and diffusion， and detailed security analysis was conducted， thereby guaranteeing the security of image transmission. Experimental results show that compared with the classical DCS method CSNet+， the proposed network achieves an average increase of 0.606 dB （0.25-1.42 dB） in Peak Signal-to-Noise Ratio （PSNR） and an average 1.11 percentage point （0.69-2.17 percentage point） improvement in Structural Similarity Index Measure （SSIM）.