<p>JPEG reversible data hiding (RDH) is a novel technique that enables hiding secret data into JPEG images by minimal modifications in original image, while ensuring perfect restoration of the original JPEG image after correctly extracting secret data. However, due to directly modifying quantized DCT coefficients, existing DCT coefficients histogram shifting-based RDH schemes often fail to fully exploit the correlation between coefficient blocks, resulting in low histogram peaks. As a result, it is very difficult for achieving an efficient balance among large embedding capacity, high visual quality, and low file size increment. To address these issues, we propose a novel solution to break the barriers in performance improvement of JPEG RDH technique. A key innovation is the design of deep neural network-based coefficient predictor, which can significantly lower the prediction error of quantized DCT coefficients to guide data embedding. To be specific, we introduce a DCT coefficient predictor based on a U-Net architecture and adopt a DCT matrix partitioning strategy to improve prediction accuracy. This combination can fully exploit the inter-block correlation of quantized DCT coefficients to further enhance the prediction error histogram, which results in embedding capacity increasing and optimizes the trade-off both the selection of embedding positions and the capacity-distortion. A series of comprehensive experiments demonstrate that the proposed scheme can obtain a superior overall performance, and efficiently achieve a better balance in terms of embedding capacity, visual quality, and file size preservation compared to existing excellent JPEG RDH schemes.</p>