{"paper":{"title":"Backdoor Channels Hidden in Latent Space: Cryptographic Undetectability in Modern Neural Networks","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Neural networks can hide backdoors as statistically indistinguishable latent directions, reducing detection to an intractable hypothesis test on model parameters.","cross_cats":["cs.LG"],"primary_cat":"cs.CR","authors_text":"Eirik Reiestad, Inga Str\\\"umke, Kristian Gj{\\o}steen, Marte Eggen","submitted_at":"2026-05-13T09:06:25Z","abstract_excerpt":"Recent cryptographic results establish that neural networks can be backdoored such that no efficient algorithm can distinguish them from a clean model. These guarantees, however, have been confined to stylised architectures of limited practical relevance, leaving open whether comparable undetectability extends to modern, end-to-end trained networks. We construct such an attack mechanism for state-of-the-art architectures, closely aligned to the cryptographic notion of undetectability, by identifying backdoor channels as learned latent directions, and show that the question of undetectability r"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"if exploitable channels within a network's latent space are statistically indistinguishable from naturally learned directions, an attacker need not introduce foreign structure but can instead exploit the geometry the network already possesses.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The hypothesis test between clean and backdoored parameter distributions is intractable in practice for state-of-the-art models; this is stated as a conjecture without a formal reduction or hardness proof.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Backdoors can be realized as statistically natural latent directions in modern neural networks, achieving high attack success with negligible clean accuracy loss and resisting existing defenses.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Neural networks can hide backdoors as statistically indistinguishable latent directions, reducing detection to an intractable hypothesis test on model parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c2aa25b68bb6af23ecae77c67ebf294b48367790eab7a4a030685130825084c3"},"source":{"id":"2605.13214","kind":"arxiv","version":1},"verdict":{"id":"3f60653d-dec8-485b-9f8e-26f5dcf66494","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:33:33.963434Z","strongest_claim":"if exploitable channels within a network's latent space are statistically indistinguishable from naturally learned directions, an attacker need not introduce foreign structure but can instead exploit the geometry the network already possesses.","one_line_summary":"Backdoors can be realized as statistically natural latent directions in modern neural networks, achieving high attack success with negligible clean accuracy loss and resisting existing defenses.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The hypothesis test between clean and backdoored parameter distributions is intractable in practice for state-of-the-art models; this is stated as a conjecture without a formal reduction or hardness proof.","pith_extraction_headline":"Neural networks can hide backdoors as statistically indistinguishable latent directions, reducing detection to an intractable hypothesis test on model parameters."},"references":{"count":31,"sample":[{"doi":"","year":2025,"title":"Backdoor attacks and defenses in computer vision domain: A survey.arXiv preprint arXiv:2509.07504, 2025","work_id":"22a13bec-74e2-44d2-8484-6385db6a5ba7","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Complexity theoretic lower bounds for sparse principal component detection","work_id":"8ace9111-4efc-425d-b55e-bd1030ac543d","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2013,"title":"Computational Lower Bounds for Sparse PCA","work_id":"c8902f8d-0de9-4766-b0d4-e00ae72c6596","ref_index":3,"cited_arxiv_id":"1304.0828","is_internal_anchor":true},{"doi":"","year":2019,"title":"Brennan and Guy Bresler","work_id":"a0e8a45c-e64c-46e6-a89b-916d0bcbe2c8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Data free backdoor attacks.Advances in Neural Information Processing Systems, 37:23881–23911, 2024","work_id":"052bbb93-8ee7-40ae-bc00-41a07f3cfd58","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":31,"snapshot_sha256":"9f6fcd43dee3685a02c3078d870d16706801aa608511822ee66ded0fc2acf78d","internal_anchors":5},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}