Abstract: Existing Robust Reversible Watermarking (RRW) methods for JPEG images often rely on a single statistical feature, fixed frequency bands, or empirically selected carrier coefficients, making it difficult to simultaneously balance robustness, imperceptibility, and reversibility under multiple types of attacks. To address this problem, an RRW method for JPEG images combining global-local statistical features with attack-driven carrier selection was proposed. The method performs watermark embedding and extraction directly in the quantized Discrete Cosine Transform (DCT) domain. First, the changes of quantized DCT coefficients under multiple attack conditions are used to construct robustness scores for candidate positions, and a lightweight Convolutional Neural Network (CNN) is employed to generate a unified robust index sequence for carrier-coefficient selection. Second, mean and median statistics are combined to construct a block-level robust feature, thereby improving its stability against global shifts and local abnormal perturbations. Finally, four-band joint symmetric modulation and feature histogram shifting are employed to achieve watermark embedding, blind extraction, and exact recovery of the original quantized DCT coefficients in the absence of attacks. Experimental results show that, with a payload of 128 bits and a JPEG Quality Factor (QF) of 90, the proposed method achieves an average Peak Signal-to-Noise Ratio (PSNR) of 41.97 dB, an average Structural Similarity Index Measure (SSIM) of 0.9749, and an average file growth rate of 1.7610% on five standard test images, while exactly recovering the original quantized DCT coefficients. Under JPEG recompression, JPEG2000 compression, WebP compression, noise, and filtering attacks, the proposed method maintains low Bit Error Rates (BERs), demonstrating good robustness and reversibility.

Key words: Keywords: Robust Reversible Watermarking (RRW), JPEG image, quantized Discrete Cosine Transform (DCT) coefficient, global-local statistical feature, attack-driven selection, histogram shifting

摘要： 针对现有JPEG图像鲁棒可逆水印（RRW）方法多依赖单一统计特征、固定频带或经验式载体系数选择，难以在多类攻击下同时兼顾鲁棒性、不可感知性和可逆性的问题，提出一种融合全局-局部统计特征及攻击驱动载体选择的JPEG图像RRW方法。该方法直接在JPEG量化离散余弦变换（DCT）系数域中进行水印嵌入与提取：首先，利用多攻击条件下量化DCT系数的变化构造候选位置鲁棒性评分，并通过轻量卷积神经网络（CNN）生成统一通用鲁棒索引序列，以指导嵌入载体系数选择；其次，结合均值统计和中值统计构造块级鲁棒特征，以增强特征对整体漂移和局部异常扰动的稳定性；最后，通过四频带联合对称调制与特征直方图平移实现水印嵌入、盲提取和无攻击条件下的原始量化DCT系数精确恢复。实验结果表明，在嵌入容量为128 bit、JPEG质量因子（QF）为90时，所提方法在5幅经典测试图像上的平均峰值信噪比（PSNR）为41.97 dB，平均结构相似性（SSIM）为0.9749，平均文件增长率为1.7610%，且均能实现原始量化DCT系数的精确恢复；在JPEG重压缩、JPEG2000压缩、WebP压缩、噪声和滤波等攻击下，所提方法保持较低误比特率，表现出良好的鲁棒性和可逆性。

关键词: 鲁棒可逆水印, JPEG图像, 量化DCT系数, 全局-局部统计特征, 攻击驱动选择, 直方图平移