The Measurement Framework for Underestimated Failure Modes of Past-only Query Policies in Open-World Malware Family Classification

Jiali Che; Haoqing Ren; Tingting Sun

doi:10.54097/evhnkb11

The Measurement Framework for Underestimated Failure Modes of Past-only Query Policies in Open-World Malware Family Classification

Jiali Che, Haoqing Ren, Tingting Sun

Frontiers in Computing and Intelligent Systems, (2026), Vol.18, No.2, pp.50-54

Published: October 29, 2026

DOI: 10.54097/evhnkb11

PDF: Download Full Text PDF

Abstract

Laboratories that evaluate open-world malware family classifiers typically ship a query or abstention policy chosen on a historical proxy window. The work argues that this past-only selection step constitutes an underestimated failure mode: the proxy-optimal policy can be unsafe on the next unseen-family window even when the candidate grid is fixed. This problem differs from training a better classifier, per-sample rejection, and learning a detection-and-labeling policy with future adaptation. It presents a rigorous measurement framework: a frozen past-only protocol, a ledger of negative selectors, split safety and efficacy estimands, and one registered instance named MPSR-RSG that never reads current future labels. On development windows covering BODMAS PE and EMBER2024 with 18 comparisons, the instance exhibits zero failures. Across 12 baseline-failure windows, its hierarchical-bootstrap 95% CI for failure-rate reduction is [0.25, 1.00].For the preregistered held-out efficacy population consisting of Avast-CTU CAPEv2 and MOTIF with 9 baseline-failure windows, the corresponding CI is [0.22, 1.00].BODMAS Phase 0a cutoffs are retained within the safety monitor rather than the confirmatory efficacy population. This arrangement follows considerations of source role, chain integrity, and estimand identifiability, rather than an attempt to improve the headline interval by excluding these data. The core claim supported by experiments is that past-only selection requires high-priority evaluation under these constraints, and the work does not prove the existence of a universal selector.

Keywords: Malware Family Classification, Open-world Evaluation, Selective Classification, Temporal Evaluation

APA Citation: Jiali Che, Haoqing Ren, Tingting Sun (2026). The Measurement Framework for Underestimated Failure Modes of Past-only Query Policies in Open-World Malware Family Classification. Frontiers in Computing and Intelligent Systems, 18(2), 50-54. https://doi.org/10.54097/evhnkb11

References

Arp, D., Quiring, E., Pendlebury, F., et al. (2022). Dos and don'ts of machine learning in computer security. In Proceedings of the 31st USENIX Security Symposium (pp. 3971–3988). USENIX Association.
Pendlebury, F., Pierazzi, F., Jordaney, R., et al. (2019). TESSERACT: Eliminating experimental bias in malware classification across space and time. In Proceedings of the 28th USENIX Security Symposium (pp. 729–746). USENIX Association.
Jordaney, R., Sharad, K., Dash, S. K., et al. (2017). Transcend: Detecting concept drift in malware classification models. In Proceedings of the 26th USENIX Security Symposium (pp. 625–642). USENIX Association.
Yang, L., Ciptadi, A., Laziuk, I., et al. (2021). BODMAS: An open dataset for learning based temporal analysis of PE malware. In 2021 IEEE Symposium on Security and Privacy Workshops (SPW) (pp. 78–84). IEEE.
Joyce, R. J., Miller, G., Roth, P., et al. (2025). EMBER2024: A benchmark dataset for holistic evaluation of malware classifiers. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5516–5526). ACM.
Zola, F., Bruse, J. L., Galar, M., et al. (2023). Temporal analysis of distribution shifts in malware classification for digital forensics. In 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) (pp. 439–450). IEEE.
Joyce, R. J., Amlani, D., Nicholas, C., et al. (2023). MOTIF: A malware reference dataset with ground truth family labels. Computers & Security, 124, 102921. https://doi.org/10.1016/j.cose.2022.102921.
Bošanský, B., Kouba, D., Manhal, O., et al. (2022). Avast-CTU public CAPE dataset. arXiv:2209.03188.
Wu, X., Guo, W., Yan, J., et al. (2023). From grim reality to practical solution: Malware classification in real-world noise. In 2023 IEEE Symposium on Security and Privacy (SP) (pp. 2602–2619). IEEE.
Barbero, F., Pendlebury, F., Pierazzi, F., et al. (2022). Transcending TRANSCEND: Revisiting malware classification in the presence of concept drift. In 2022 IEEE Symposium on Security and Privacy (SP) (pp. 805–823). IEEE.
McFadden, S., Foley, M., D'Onghia, M., et al. (2026). DRMD: Deep reinforcement learning for malware detection under concept drift. In Proceedings of the AAAI Conference on Artificial Intelligence, 40(2), 854–862.
Herzog, A., Eusebi, A., Cavallaro, L. (2025). On the reliability and stability of selective methods in malware classification tasks. arXiv:2505.22843.
Vaze, S., Han, K., Vedaldi, A., et al. (2022). Open-set recognition: A good closed-set classifier is all you need. In Proceedings of the International Conference on Learning Representations (ICLR).
Lu, T., & Wang, J. (2024). DOMR: Toward deep open-world malware recognition. IEEE Transactions on Information Forensics and Security, 19, 1455–1468. https://doi.org/10.1109/TIFS.2024.3357040.
Hendrycks, D., & Gimpel, K. (2017). A baseline for detecting misclassified and out-of-distribution examples in neural networks. In Proceedings of the International Conference on Learning Representations (ICLR).
Geifman, Y., & El-Yaniv, R. (2017). Selective classification for deep neural networks. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017) (pp. 1–12). Curran Associates.
Gulrajani, I., & Lopez-Paz, D. (2021). In search of lost domain generalization. In Proceedings of the International Conference on Learning Representations (ICLR).
Ye, H., Xie, C., Cai, T., et al. (2021). Towards a theoretical framework of out-of-distribution generalization. In Advances in Neural Information Processing Systems 34 (NeurIPS 2021) (pp. 24608–24621). Curran Associates.