近年生成式人工智慧（Generative AI）影像技術快速發展，文字生成影像模型（text-to-image models）已能產生高度逼真的視覺內容，並廣泛影響數位內容生態，使數位影像真偽鑑識與內容信任面臨更複雜的挑戰。既有整張影像偵測模型（full-image detectors）雖可取得良好的評估效能，但較難說明其判斷所依賴的局部證據（local evidence），也不易量化推論過程所需的證據預算（evidence budget），對其在壓縮、轉傳與再數位化等真實流通條件下的失效行為與邊界解釋也相對有限。為回應此問題，本文以 RRDataset / RRBench 的真實退化條件為基礎，提出「計算感知局部證據選擇」（Compute-Aware Evidence Selection, CAES）分析框架。此框架串接候選區塊建構、特徵排序與隨時可停式（anytime）聚合機制，以追蹤局部線索的可用性並量化局部證據需求。
實驗結果顯示，局部證據的可用性具有明顯的情境相依性。在未經退化的原始情境中，加入提前停止機制的影像區塊分支（Patch Early Stopping, Patch ES）呈現較強的鑑識能力，AUC 達 0.9532，並明顯高於匹配式整張影像基準模型；相對地，再數位化流程會削弱高頻與局部殘差特徵，使影像區塊分支效能降至 AUC 0.6424，凸顯局部鑑識線索在此情境下的失效邊界。
進一步的分數層級互補性分析（CAES-G）顯示，在固定測試切分中結合局部敏感度與整體影像脈絡可形成互補的診斷訊號，整體 AUC 達 0.8875。本文將此結果作為事後分數層級後融合診斷，用以觀察局部與整體證據在偵測任務中的互補關係。本論文因此將局部證據的動態累積、證據預算權衡與真實世界失效邊界納入同一分析視角，為後續 AI 生成影像偵測方法的證據分析提供受控實證依據。

Recent advances in generative artificial intelligence have enabled text-to-image models to produce highly realistic visual content, making them increasingly common in digital media and creating more complex challenges for digital image forensics and content trust. Although full-image detectors can achieve strong evaluation performance, they provide limited visibility into the local evidence on which decisions depend, the evidence budget required during inference, and the failure patterns and boundaries that emerge under real-world circulation conditions such as compression, forwarding, and re-digitization. To examine these issues, this thesis proposes Compute-Aware Evidence Selection (CAES), an analysis framework grounded in the real-degradation conditions of RRDataset / RRBench. This framework integrates candidate-region construction, feature-based ranking, and anytime aggregation to track the availability of local cues and quantify local evidence requirements.
The results indicate that the usefulness of local evidence is strongly scenario-dependent. In the original, non-degraded scenario, the early-stopping patch branch (Patch ES) shows strong forensic capability, reaching an AUC of 0.9532 and outperforming the matched full-image baseline. In contrast, re-digitization weakens high-frequency and local residual cues, reducing patch-branch performance to an AUC of 0.6424 and highlighting a failure boundary for local forensic evidence under this scenario.
A further score-level complementarity analysis (CAES-G) shows that combining local sensitivity with global context produces a complementary diagnostic signal on the fixed test split, achieving an overall AUC of 0.8875. This result is used as a post-hoc score-level late-fusion diagnostic that supports the complementary value of local and global evidence in the detection task. Overall, this study brings the dynamic accumulation of local evidence, evidence-budget trade-offs, and real-world failure boundaries into one analytical perspective, providing controlled empirical grounding for future evidence analysis in AI-generated image detection.

[1] L. Verdoliva, “Media forensics and deepfakes: An overview,” IEEE J. Sel. Topics Signal Process., vol. 14, no. 5, pp. 910–932, 2020. https://doi.org/10.1109/JSTSP.2020.3002101
[2] S.-Y. Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, “CNN-generated images are surprisingly easy to spot... for now,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 8692–8701, 2020. https://doi.org/10.1109/CVPR42600.2020.00872
[3] C. Li et al., “Bridging the gap between ideal and real-world evaluation: Benchmarking AI-generated image detection in challenging scenarios,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 20379–20389, 2025. https://doi.org/10.1109/ICCV51701.2025.01895
[4] P. Fernandez, G. Couairon, H. Jégou, M. Douze, and T. Furon, “The Stable Signature: Rooting watermarks in latent diffusion models,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 22409–22420, 2023. https://doi.org/10.1109/ICCV51070.2023.02053
[5] K. Balan, S. Agarwal, S. Jenni, A. Parsons, A. Gilbert, and J. Collomosse, “EKILA: Synthetic media provenance and attribution for generative art,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), pp. 913–922, 2023. https://doi.org/10.1109/CVPRW59228.2023.00098
[6] M. Guo et al., “AI-generated image detection: Passive or watermark?” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), pp. 400–410, 2026. https://openaccess.thecvf.com/content/CVPR2026W/SAFE/html/Guo_AI-generated_Image_Detection_Passive_or_Watermark_CVPRW_2026_paper.html
[7] J. Frank, T. Eisenhofer, L. Schönherr, A. Fischer, D. Kolossa, and T. Holz, “Leveraging frequency analysis for deep fake image recognition,” in Proc. 37th Int. Conf. Mach. Learn. (ICML), Proc. Mach. Learn. Res., vol. 119, pp. 3247–3258, 2020. https://proceedings.mlr.press/v119/frank20a.html
[8] H. Farid, “Image forgery detection,” IEEE Signal Process. Mag., vol. 26, no. 2, pp. 16–25, 2009. https://doi.org/10.1109/MSP.2008.931079
[9] A. Piva, “An overview on image forensics,” ISRN Signal Process., vol. 2013, Art. no. 496701, 22 pp., 2013. https://doi.org/10.1155/2013/496701
[10] L. Yuan, X. Li, Y. Zhang, J. Zhang, H. Li, and X. Gao, “MLEP: Multi-granularity local entropy patterns for generalized AI-generated image detection,” in Adv. Neural Inf. Process. Syst., vol. 38, pp. 68981–69000, 2025. https://papers.nips.cc/paper_files/paper/2025/hash/63db94e06ce85dde10871c99dd39ad98-Abstract-Conference.html
[11] S. Liang, J. Liu, R. Chen, and Q. Guan, “FerretNet: Efficient synthetic image detection via local pixel dependencies,” in Adv. Neural Inf. Process. Syst., vol. 38, pp. 42109–42133, 2025. https://papers.nips.cc/paper_files/paper/2025/hash/3bfcc45018b4156b92197db47e9742bd-Abstract-Conference.html
[12] H. Liu, Z. Tan, C. Tan, Y. Wei, J. Wang, and Y. Zhao, “Forgery-aware adaptive transformer for generalizable synthetic image detection,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 10770–10780, 2024. https://doi.org/10.1109/CVPR52733.2024.01024
[13] Z. Yang et al., “All patches matter, more patches better: Enhance AI-generated image detection via panoptic patch learning,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2026. https://iclr.cc/virtual/2026/poster/10007395
[14] H.-S. Chen, S. Hu, S. You, and C.-C. J. Kuo, “DefakeHop++: An enhanced lightweight deepfake detector,” APSIPA Trans. Signal Inf. Process., vol. 11, no. 2, pp. 1–21, 2022. https://doi.org/10.1561/116.00000126
[15] A. Howard et al., “Searching for MobileNetV3,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 1314–1324, 2019. https://doi.org/10.1109/ICCV.2019.00140
[16] A. Gushchin et al., “NTIRE 2026 challenge on robust AI-generated image detection in the wild,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), pp. 1895–1913, 2026. https://openaccess.thecvf.com/content/CVPR2026W/NTIRE/html/Gushchin_NTIRE_2026_Challenge_on_Robust_AI-Generated_Image_Detection_in_the_CVPRW_2026_paper.html
[17] C. Tan et al., “Rethinking the up-sampling operations in CNN-based generative network for generalizable deepfake detection,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 28130–28139, 2024. https://doi.org/10.1109/CVPR52733.2024.02657
[18] Z. Wang et al., “DIRE for diffusion-generated image detection,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pp. 22388–22398, 2023. https://doi.org/10.1109/ICCV51070.2023.02051
[19] J. Ricker, D. Lukovnikov, and A. Fischer, “AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 9130–9140, 2024. https://doi.org/10.1109/CVPR52733.2024.00872
[20] Y. Ju, S. Jia, L. Ke, H. Xue, K. Nagano, and S. Lyu, “Fusing global and local features for generalized AI-synthesized image detection,” in Proc. IEEE Int. Conf. Image Process. (ICIP), pp. 3465–3469, 2022. https://doi.org/10.1109/ICIP46576.2022.9897820
[21] Y. Ju, S. Jia, J. Cai, H. Guan, and S. Lyu, “GLFF: Global and local feature fusion for AI-synthesized image detection,” IEEE Trans. Multimedia, vol. 26, pp. 4073–4085, 2024. https://doi.org/10.1109/TMM.2023.3313503
[22] U. Ojha, Y. Li, and Y. J. Lee, “Towards universal fake image detectors that generalize across generative models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 24480–24489, 2023. https://doi.org/10.1109/CVPR52729.2023.02345
[23] S. Yan et al., “A sanity check for AI-generated image detection,” in Proc. Int. Conf. Learn. Represent. (ICLR), pp. 70702–70720, 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/b0303773962ea1b5394c3a83cc7dd066-Abstract-Conference.html
[24] S. Teerapittayanon, B. McDanel, and H. T. Kung, “BranchyNet: Fast inference via early exiting from deep neural networks,” in Proc. 23rd Int. Conf. Pattern Recognit. (ICPR), pp. 2464–2469, 2016. https://doi.org/10.1109/ICPR.2016.7900006
[25] C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in Proc. 34th Int. Conf. Mach. Learn. (ICML), Proc. Mach. Learn. Res., vol. 70, pp. 1321–1330, 2017. https://proceedings.mlr.press/v70/guo17a.html
[26] R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green AI,” Commun. ACM, vol. 63, no. 12, pp. 54–63, 2020. https://doi.org/10.1145/3381831
[27] M. Zhu et al., “GenImage: A million-scale benchmark for detecting AI-generated image,” in Adv. Neural Inf. Process. Syst., vol. 36, Datasets and Benchmarks Track, pp. 77771–77782, 2023. https://doi.org/10.52202/075280-3398
[28] Z. Li et al., “Is artificial intelligence generated image detection a solved problem?” in Adv. Neural Inf. Process. Syst., vol. 38, Datasets and Benchmarks Track, 2025. https://papers.nips.cc/paper_files/paper/2025/hash/fb693c67f61e5321746ffce8b6fdd2d0-Abstract-Datasets_and_Benchmarks_Track.html
[29] S. Gye, J. Ko, H. Shon, M. Kwon, and J. Kim, “Reducing the content bias for AI-generated image detection,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), pp. 399–408, 2025. https://doi.org/10.1109/WACV61041.2025.00049
[30] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap. New York, NY, USA: Chapman & Hall, 1993.
[31] M. Mitchell et al., “Model cards for model reporting,” in Proc. Conf. Fairness, Accountability, Transparency (FAT*), pp. 220–229, 2019. https://doi.org/10.1145/3287560.3287596
[32] R. R. Selvaraju et al., “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), pp. 618–626, 2017. https://doi.org/10.1109/ICCV.2017.74
[33] Z. Ye and F. Farnia, “Gaussian smoothing in saliency maps: The stability-fidelity trade-off in neural network interpretability,” in Proc. 28th Int. Conf. Artif. Intell. Stat. (AISTATS), Proc. Mach. Learn. Res., vol. 258, pp. 2125–2133, 2025. https://proceedings.mlr.press/v258/ye25a.html
[34] V. Petsiuk, A. Das, and K. Saenko, “RISE: Randomized input sampling for explanation of black-box models,” in Proc. Brit. Mach. Vis. Conf. (BMVC), paper no. 151, pp. 1–13, 2018. https://bmva-archive.org.uk/bmvc/2018/contents/papers/1064.pdf
[35] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in Computer Vision—ECCV 2014, Part I, Lecture Notes in Computer Science, vol. 8689, pp. 818–833, 2014. https://doi.org/10.1007/978-3-319-10590-1_53