Video steganography is an important technique for secure information transmission by concealing secret information within video carriers. With the rapid development of social networks, however, video steganography faces significant challenges from heavy recompression performed by social platforms to improve transmission efficiency, as well as cropping attacks caused by unauthorized reposting of videos. To address the problem that videos transmitted over social networks are vulnerable to severe recompression and border-cropping attacks, a robust video steganography algorithm based on face detection is proposed. First, a Haar cascade detector is employed to locate the facial regions of interest (ROIs), followed by block alignment and overlapping-region merging of the ROIs. Then, block texture features are extracted from the Y component using three-level DTCWT-PCA, while secret information is embedded into the Haar-DWT/SVD domain of the U component by combining adaptive quantization index modulation, syndrome-trellis coding (STC), and Reed–Solomon (RS) error correction coding, so as to balance robustness and distortion control. Experimental results show that the proposed algorithm achieves an average bit error rate of 1.55% under eight types of local heavy recompression channels. Under border-cropping attacks, provided that the ROI remains intact, the average bit error rate is 1.62%. The average PSNR and SSIM are 41.13 dB and 0.996, respectively. Compared with MEC-AQIM, the proposed method maintains comparable robustness against heavy recompression while further exhibiting resistance to border-cropping attacks.