A Lightweight Multi-Branch Network for Double JPEG Compression Detection

Double JPEG compression detection is a key forensic tool for identifying image tampering, yet existing convolutional architectures often achieve robustness to real-world quantization-table diversity at the cost of increased model size. Recent approaches improve robustness by providing the quantization table as an additional network input, while multi-branch architectures that allocate an independent convolutional branch to each of the first 27 AC frequency bands achieve high detection performance at the cost of increased model complexity. This paper proposes a lightweight variant of this multi-branch architecture that replaces the per-band branches with a single weight-shared encoder, a learnable band embedding, and multi-query attention pooling, reducing the number of encoder parameters 27-fold and the total parameter count to 9.85M while retaining the inter-branch DenseNet-121 backbone and the quantization-table injection strategy. Using 1095 real-world quantization tables grouped into five equivalent quality-factor pools, the proposed model is evaluated for aligned double compression with different quantization tables, for both patch-level classification and pixel-level manipulation localization on the McGill and JPEG AI CTTC datasets. An additional variant without quantization-table input is trained to examine the contribution of quantization-table information to detection performance. On test images independent of the training data, the proposed method achieves the highest average patch-level F1 score among the compared models, while on the validation set, its accuracy matches that of the reference architecture within the variation measured across training seeds for the proposed model, with only approximately 22% of the parameters of the reference architecture. For pixel-level localization, the proposed method achieves an F1 score of 0.580, compared with 0.577 and 0.564 for the reference architectures. Removing the quantization-table input decreases the patch-level classification accuracy on the validation set from 94.16% to 91.23%, indicating that the proposed model remains applicable when this information is unavailable.