Безпека ланцюга поставок ML і підписання моделі

An ML system depends on more than model weights. Its supply chain includes source code, build tools, package dependencies, base images, training and evaluation data, serialized artifacts and deployment configuration. An attacker or accidental error can introduce a malicious dependency, alter a model file, leak credentials during a build or deploy an artifact that was never reviewed. Security practices should protect these inputs and preserve traceable evidence about how release artifacts were produced. Digital signatures bind an artifact to a signing identity and a cryptographic digest. A verifier can check that the artifact matches what was signed and that the signature chains to an expected identity or policy. Provenance statements can record build inputs and process details. Sigstore tooling supports signing container artifacts and verification workflows. These checks help reject tampered or unauthorized artifacts, but a valid signature only says that a particular identity signed particular bytes; it does not mean the contents are safe or correct. Model serialization deserves special care. Python pickle can execute arbitrary code during deserialization, so loading an untrusted pickle is dangerous. Verify artifacts from trusted sources and consider formats with narrower execution behavior when appropriate. Even then, data parsers and model runtimes can have vulnerabilities. Scan dependencies and containers, restrict permissions, isolate training and inference, and avoid storing secrets in images or model bundles. A useful release record links source revision, build environment, dependencies, data lineage, evaluation report and artifact digest. Apply access control to signing keys and prefer short-lived identities or managed signing workflows where available. Verify signatures in deployment policy rather than signing without enforcement. Security review should include incident response and key rotation. Supply-chain controls establish integrity and provenance; separate testing, privacy and responsible-AI evaluations are still needed to judge model behavior and fitness for use.