Securing the open-source ecosystem: AI-enhanced and explainable supply chain security for reliable software development

Abstract

Modern software supply chains face increasing risks from vulnerable and anomalous dependencies, necessitating automated and interpretable detection methods integrated within Continuous Integration and Continuous Deployment (CI/CD) workflows. Open-source software (OSS) ecosystems are increasingly targeted by sophisticated attacks—including AI model poisoning, dependency confusion, contributor hijacking, and transitive dependency anomalies—that often evade traditional CVE-centric tools such as Snyk and Dependabot. These reactive solutions are primarily designed for known vulnerabilities and may exhibit high false-positive rates when applied to behavioral anomalies, while also lacking interpretable and actionable guidance for developers. This limitation creates a critical barrier to reliable and secure software development in modern DevSecOps environments. To address this gap, this paper introduces XAI-SCS, an intelligent, explainable decision-support framework for OSS supply chain security that performs metadata-driven behavioral risk assessment to proactively identify anomalous package activity. A hybrid CNN–LSTM–Autoencoder architecture achieves strong anomaly-detection performance on the held-out test set (F1 = 0.91), with low repeated-run variability (mean F1 = 0.90 ± 0.02 across five runs) on a federated dataset of 12,000 npm/PyPI package versions, primarily trained using CVE-linked labels and GAN-simulated poisoning samples. A separate held-out evaluation set of 240 manually curated real-world non-CVE OSS supply-chain incidents was used exclusively for external generalization assessment. Accordingly, the non-CVE results are interpreted as preliminary evidence of transferability to selected metadata-visible behavioral anomalies rather than comprehensive real-world non-CVE threat coverage. To support transparent developer decision-making, SHAP-based explanations and counterfactual remediation suggestions produced high perceived clarity (4.3 ± 0.4/5) and actionability (4.1 ± 0.5/5) and self-reported remediation intent (96%, n  = 82). However, these findings reflect developer perceptions in a controlled study and do not yet demonstrate measured improvements in real-world remediation time, fix completion, or vulnerability resolution within operational DevSecOps teams. Optional high-assurance deployment extensions, including Federated Continual Learning, zero-knowledge attestation via zk-SNARKs, and post-quantum secure enclaves, are presented as exploratory architectural pathways for future privacy-preserving, verifiable, and confidential deployment, rather than as part of the core empirical evaluation. Embedded in controlled, CI/CD-representative workflows with 0.8-second average latency, XAI-SCS remains competitive with CVE-centric tools on known-vulnerability cases while providing preliminary evidence of complementary metadata-level behavioral risk screening in GAN-simulated and limited curated non-CVE evaluation subsets. The comparison is therefore interpreted as evidence that XAI-SCS can complement CVE-centric SCA tools by adding metadata-level behavioral screening and developer-facing explanations, rather than replacing or universally outperforming tools such as Snyk and Dependabot. This work’s primary contribution is a metadata-driven, CI/CD-integrated intelligent decision-support framework that delivers high-accuracy behavioral anomaly detection (F1 = 0.91) and transparent, developer-facing explanations (clarity 4.3/5, remediation intent 96%) for OSS supply chain risk assessment. Optional high-assurance deployment extensions (federated learning, cryptographic attestation, secure enclaves) are discussed separately and do not form part of the core empirical evaluation.