Memulihkan Reproduksibilitas Bit-Eksak pada Benchmark Steganalisis di bawah Kompresi JPEG Berantai dan Derau Gauss melalui Pemulihan Salt Hash dan Audit Artefak Pelatihan
DOI:
https://doi.org/10.26740/jte.v15n3.p313-327Keywords:
steganalysis benchmark, reproducibility, DataLogger, hash randomization, Adaptive LSB-MMEDAbstract
Klaim ketahanan untuk steganalyzer deep learning hanya dapat dipercaya sejauh citra benchmark tempat klaim itu diukur dapat dipercaya, padahal dataset steganalisis jarang diperiksa dari segi regenerasi bit-eksak, konsistensi artefak pelatihan, maupun statistik yang dapat dihitung ulang. Penelitian ini membangun benchmark berpasangan dan terpisah per sumber (source-disjoint) dengan 13.000 citra (2.500 sumber BOSSbase, Adaptive LSB-MMED pada 0,2 bpp, sembilan kondisi JPEG dan derau Gauss tunggal maupun berantai), dan merilisnya bersama manifes SHA-256, metadata Croissant 1.0 yang tervalidasi, sembilan eksekusi detektor baseline, serta protokol reproduksibilitas empat tingkat yang mencakup integritas berkas, regenerasi bit-eksak, audit artefak pelatihan, dan penghitungan ulang metrik. Seluruh 12.200 berkas yang diaudit cocok dengan manifes, dan pembagian data, seluruh 2.100 hasil penyisipan yang tersedia, serta kedua kondisi JPEG teregenerasi identik byte demi byte. Sebaliknya, keenam kondisi derau meregenerasi 0% berkasnya pada setiap pengaturan interpreter, karena seed derau bergantung pada hash string Python yang diacak per proses. Pencarian menyeluruh atas 65.536 kemungkinan salt 16-bit mengidentifikasi satu salt unik per kondisi hanya dari lima sampai tujuh piksel, tidak cocok dengan nilai PYTHONHASHSEED mana pun hingga 1,2 × 107, dan memulihkan regenerasi identik byte demi byte atas seluruh 6.000 berkas ber-derau. Realisasi derau alternatif mengubah AUC detektor terbaik dalam rentang 0,010. Audit artefak meloloskan sembilan eksekusi yang dirilis dan mendeteksi keempat kesalahan yang disuntikkan, sedangkan penghitungan ulang metrik dari prediksi per citra mengoreksi kesalahan agregasi satu-seed dan uji DeLong yang mengabaikan kovarians, sehingga membalik satu keputusan signifikansi. Oleh karena itu, seed tetap dan kode yang dipublikasikan tidak cukup; seed harus menghindari hash ber-salt, dan regenerasi harus diverifikasi pada tingkat byte.
Downloads
References
Akhtar, M. dkk. 2024. Croissant: A metadata format for ML-ready datasets. Dalam: Advances in Neural Information Processing Systems 37 (NeurIPS 2024), Datasets and Benchmarks Track, hlm. 82133–82148. doi: 10.52202/079017-2610.
Alrusaini, O. 2025. Deep learning for steganalysis: Evaluating model robustness against image transformations. Front. Artif. Intell., 8: Art. no. 1532895. doi: 10.3389/frai.2025.1532895.
Aumasson, J.-P. dan Bernstein, D. J. 2012. SipHash: A fast short-input PRF. Dalam: Progress in Cryptology – INDOCRYPT 2012, LNCS, vol. 7668. Berlin, Germany: Springer, hlm. 489–508. doi: 10.1007/978-3-642-34931-7_28.
Bas, P., Filler, T., dan Pevný, T. 2011. ‘Break our steganographic system’: The ins and outs of organizing BOSS. Dalam: Information Hiding (IH 2011), LNCS, vol. 6958. Berlin, Germany: Springer, hlm. 59–70. doi: 10.1007/978-3-642-24178-9_5.
Boroumand, M., Chen, M., dan Fridrich, J. 2019. Deep residual network for steganalysis of digital images. IEEE Trans. Inf. Forensics Security, 14(5): 1181–1193. doi: 10.1109/TIFS.2018.2871749.
Bouthillier, X. dkk. 2021. Accounting for variance in machine learning benchmarks. Dalam: Proc. Mach. Learn. Syst. (MLSys), vol. 3, hlm. 747–769.
Chakraborty, S., Jalal, A. S., dan Bhatnagar, C. 2017. LSB based non blind predictive edge adaptive image steganography. Multimedia Tools Appl., 76(6): 7973–7987. doi: 10.1007/s11042-016-3449-4.
Cogranne, R., Giboulot, Q., dan Bas, P. 2020. ALASKA#2: Challenging academic research on steganalysis with realistic images. Dalam: Proc. IEEE Int. Workshop on Information Forensics and Security (WIFS), New York, NY, USA, Dec. 2020, hlm. 1–5. doi: 10.1109/WIFS49906.2020.9360896.
DeLong, E. R., DeLong, D. M., dan Clarke-Pearson, D. L. 1988. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics, 44(3): 837–845. doi: 10.2307/2531595.
Efron, B. dan Tibshirani, R. J. 1993. An Introduction to the Bootstrap. New York, NY, USA: Chapman & Hall.
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumé III, H., dan Crawford, K. 2021. Datasheets for datasets. Commun. ACM, 64(12): 86–92. doi: 10.1145/3458723.
Giboulot, Q., Cogranne, R., Borghys, D., dan Bas, P. 2020. Effects and solutions of cover-source mismatch in image steganalysis. Signal Process.: Image Commun., 86: Art. no. 115888. doi: 10.1016/j.image.2020.115888.
Gundersen, O. E., Coakley, K., Kirkpatrick, C., dan Gil, Y. 2022. Sources of irreproducibility in machine learning: A review. arXiv:2204.07610. Tersedia pada: https://arxiv.org/abs/2204.07610
Gundersen, O. E. dan Kjensmo, S. 2018. State of the art: Reproducibility in artificial intelligence. Dalam: Proc. AAAI Conf. Artif. Intell., vol. 32, no. 1. doi: 10.1609/aaai.v32i1.11503.
Heimes, C. 2013. PEP 456 – Secure and interchangeable hash algorithm. Python Enhancement Proposals. Tersedia pada: https://peps.python.org/pep-0456/
Holm, S. 1979. A simple sequentially rejective multiple test procedure. Scand. J. Statist., 6(2): 65–70. Tersedia pada: https://www.jstor.org/stable/4615733
Kapoor, S. dan Narayanan, A. 2023. Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9): Art. no. 100804. doi: 10.1016/j.patter.2023.100804.
Ker, A. D., Bas, P., Böhme, R., Cogranne, R., Craver, S., Filler, T., Fridrich, J., dan Pevný, T. 2013. Moving steganography and steganalysis from the laboratory into the real world. Dalam: Proc. 1st ACM Workshop on Information Hiding and Multimedia Security (IH&MMSec), hlm. 45–58. doi: 10.1145/2482513.2482965.
Kheddar, H., Hemis, M., Himeur, Y., Megías, D., dan Amira, A. 2024. Deep learning for steganalysis of diverse data types: A review of methods, taxonomy, challenges and future directions. Neurocomputing, 581: Art. no. 127528. doi: 10.1016/j.neucom.2024.127528.
Kombrink, M., van Lierop, S., Stolwijk, D., Worring, M., Vrijdag, D., dan Geradts, Z. 2025. REVEAL: A large-scale comprehensive image dataset for steganalysis. Forensic Sci. Int.: Digit. Investig., 55: Art. no. 302006. doi: 10.1016/j.fsidi.2025.302006.
Lamb, C. dan Zacchiroli, S. 2022. Reproducible builds: Increasing the integrity of software supply chains. IEEE Softw., 39(2): 62–70. doi: 10.1109/MS.2021.3073045.
Liang, W. dan Li, Q. 2026. Steganalysis network for weak steganographic signal extraction and enhancement. Sensors, 26(4): Art. no. 1329. doi: 10.3390/s26041329.
Longpre, S. dkk. 2024. A large-scale audit of dataset licensing and attribution in AI. Nat. Mach. Intell., 6(8): 975–987. doi: 10.1038/s42256-024-00878-8.
Mallet, A., Beneš, M., dan Cogranne, R. 2024. Cover-source mismatch in steganalysis: Systematic review. EURASIP J. Inf. Secur., 2024: Art. no. 26. doi: 10.1186/s13635-024-00171-6.
McNemar, Q. 1947. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2): 153–157. doi: 10.1007/BF02295996.
Pham, H. V., Qian, S., Wang, J., Lutellier, T., Rosenthal, J., Tan, L., Yu, Y., dan Nagappan, N. 2020. Problems and opportunities in training deep learning software systems: An analysis of variance. Dalam: Proc. 35th IEEE/ACM Int. Conf. Automated Software Engineering (ASE), hlm. 771–783. doi: 10.1145/3324884.3416545.
Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivière, V., Beygelzimer, A., d’Alché-Buc, F., Fox, E., dan Larochelle, H. 2021. Improving reproducibility in machine learning research (A report from the NeurIPS 2019 reproducibility program). J. Mach. Learn. Res., 22(164): 1–20. Tersedia pada: https://jmlr.org/papers/v22/20-303.html
Python Software Foundation. 2022. What’s new in Python 3.11. Python 3 documentation. Tersedia pada: https://docs.python.org/3/whatsnew/3.11.html
Semmelrock, H., Ross-Hellauer, T., Kopeinik, S., Theiler, D., Haberl, A., Thalmann, S., dan Kowald, D. 2025. Reproducibility in machine-learning-based research: Overview, barriers, and drivers. AI Mag., 46(2): Art. no. e70002. doi: 10.1002/aaai.70002.
Wang, Z., Bovik, A. C., Sheikh, H. R., dan Simoncelli, E. P. 2004. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600–612. doi: 10.1109/TIP.2003.819861.
Wilkinson, M. D. dkk. 2016. The FAIR guiding principles for scientific data management and stewardship. Sci. Data, 3: Art. no. 160018. doi: 10.1038/sdata.2016.18.
Yang, W., Liu, Y., Chen, Y., Cui, Y., Guo, C., Shen, G., dan Wu, H. 2026. HSMNet: A multi-resolution grayscale image steganalysis method based on hybrid dilated convolution and self-attention multi-channel network. Inf. Sci., 728: Art. no. 122824. doi: 10.1016/j.ins.2025.122824.
Yang, Z., Wang, K., Ma, S., Huang, Y., Kang, X., dan Zhao, X. 2020. IStego100K: Large-scale image steganalysis dataset. Dalam: Digital Forensics and Watermarking (IWDW 2019), LNCS, vol. 12022. Cham, Switzerland: Springer, hlm. 352–364. doi: 10.1007/978-3-030-43575-2_29.
Yousfi, Y. dan Fridrich, J. 2020. JPEG steganalysis detectors scalable with respect to compression quality. Dalam: Proc. IS&T Electronic Imaging, Media Watermarking, Security, and Forensics, hlm. 75-1–75-11. doi: 10.2352/ISSN.2470-1173.2020.4.MWSF-075.
Zhuang, D., Zhang, X., Song, S. L., dan Hooker, S. 2022. Randomness in neural network training: Characterizing the impact of tooling. Dalam: Proc. Mach. Learn. Syst. (MLSys), vol. 4. Tersedia pada: https://proceedings.mlsys.org/paper_files/paper/2022/file/427e0e886ebf87538afdf0badb805b7f-Paper.pdf
Downloads
Published
How to Cite
Issue
Section
Abstract views: 0
,
PDF Downloads: 0
