Adversarial ML coverage for engineers shipping ML systems. Membership inference, model extraction, evasion attacks, training-data extraction, backdoors — focused on what's exploitable against deployed models and what defenders can actually do about it. PoCs against open models, behavioral analysis for closed ones.https://adversarialml.dev/enhttps://adversarialml.dev/posts/weekly-plaa-packet-level-adversarial-attacks-in-network-traffic-det/https://adversarialml.dev/posts/weekly-plaa-packet-level-adversarial-attacks-in-network-traffic-det/A new arXiv paper builds adversarial network traffic at the packet level instead of the flow level, hitting a 92.78% evasion rate against deep-learning NIDS. Here's why that framing matters more than the number.Tue, 30 Jun 2026 00:00:00 GMTadversarial-mlnidsevasion-attacksnetwork-securityfeature-space-attacksAdversarial ML Editorialhttps://adversarialml.dev/posts/testing-robustness-against-unforeseen-adversaries/https://adversarialml.dev/posts/testing-robustness-against-unforeseen-adversaries/OpenAI's Unforeseen Attack Robustness metric quantifies how well a classifier holds up against adversarial perturbations outside its training distributionSat, 13 Jun 2026 00:00:00 GMTadversarial-robustnessevaluationneural-networksunforeseen-attacksmetricsAdversarial ML Editorialhttps://adversarialml.dev/posts/how-model-inversion-attacks-work/https://adversarialml.dev/posts/how-model-inversion-attacks-work/Embedding inversion recovers the original text from a model's embedding vectors, breaking the assumption that embeddings are an opaque, privacy-safeSat, 13 Jun 2026 00:00:00 GMTembedding-inversionvec2textprivacyml-securityvector-storeadversarial-mlAdversarial ML Editorialhttps://adversarialml.dev/posts/adversarial-training-methods-trades-mart/https://adversarialml.dev/posts/adversarial-training-methods-trades-mart/Adversarial training is the most defensible empirical robustness method, but 'adversarial training' isn't one thing.Fri, 22 May 2026 00:00:00 GMTadversarial-trainingtradesmartrobustnessrobustness-accuracy-tradeoffadversarial-mlAdversarial ML Editorialhttps://adversarialml.dev/posts/evaluating-adversarial-robustness-correctly/https://adversarialml.dev/posts/evaluating-adversarial-robustness-correctly/Most defenses that claim robustness are later broken — not because the idea was bad, but because the evaluation was.Fri, 22 May 2026 00:00:00 GMTrobustness-evaluationautoattackrobustbenchadaptive-attacksgradient-maskingadversarial-mlAdversarial ML Editorialhttps://adversarialml.dev/posts/adversarial-examples-vs-data-poisoning/https://adversarialml.dev/posts/adversarial-examples-vs-data-poisoning/Adversarial examples attack a deployed model at inference; data poisoning attacks the model before it is deployed.Mon, 11 May 2026 00:00:00 GMTadversarial-examplesdata-poisoningevasion-attacksbackdoor-attacksadversarial-mlml-securitythreat-modelingAdversarial ML Editorialhttps://adversarialml.dev/posts/membership-inference-vs-model-inversion/https://adversarialml.dev/posts/membership-inference-vs-model-inversion/Membership inference asks 'was this sample in the training set?' Model inversion asks 'what samples were in the training set?Mon, 11 May 2026 00:00:00 GMTmembership-inferencemodel-inversiontraining-data-privacydifferential-privacyml-privacyadversarial-mlml-securityAdversarial ML Editorialhttps://adversarialml.dev/posts/adversarial-attacks-vision-language-models/https://adversarialml.dev/posts/adversarial-attacks-vision-language-models/Vision-language models expand the adversarial attack surface beyond image classifiers: adversarial images can manipulate text outputs, carry visualMon, 11 May 2026 00:00:00 GMTmultimodal-adversarial-attacksvision-language-modelsadversarial-mlclip-attacksllm-securityvisual-adversarial-examplesjailbreakingAdversarial ML Editorialhttps://adversarialml.dev/posts/adversarial-patch-attacks/https://adversarialml.dev/posts/adversarial-patch-attacks/Adversarial patches are large, visible, localized perturbations designed to survive physical-world conditions — printing, lighting, and camera optics.Mon, 11 May 2026 00:00:00 GMTadversarial-patchesphysical-adversarial-attacksadversarial-examplesadversarial-mlobject-detectionautonomous-vehiclesAdversarial ML Editorialhttps://adversarialml.dev/posts/universal-adversarial-perturbations/https://adversarialml.dev/posts/universal-adversarial-perturbations/Unlike per-image attacks, universal adversarial perturbations are input-agnostic: a single crafted noise vector causes misclassification across virtuallyMon, 11 May 2026 00:00:00 GMTuniversal-adversarial-perturbationsadversarial-examplesadversarial-mlinput-agnostic-attacksimage-classifiersadversarial-robustnessAdversarial ML Editorialhttps://adversarialml.dev/posts/adversarial-robustness-nlp-text/https://adversarialml.dev/posts/adversarial-robustness-nlp-text/Discrete input spaces, semantic constraints, and human-perceptibility rules change what counts as an adversarial example in text.Sun, 10 May 2026 00:00:00 GMTadversarial-nlptext-attacksrobustnessnlpadversarial-mlml-securitytransformersAdversarial ML Editorialhttps://adversarialml.dev/posts/data-poisoning-backdoor-attacks/https://adversarialml.dev/posts/data-poisoning-backdoor-attacks/Training data manipulation, backdoor triggers, and Trojan attacks against large-scale models. What the threat model actually requires and where theSun, 10 May 2026 00:00:00 GMTdata-poisoningbackdoor-attackstrojan-mladversarial-mlml-securityfoundation-modelsAdversarial ML Editorialhttps://adversarialml.dev/posts/evasion-attacks-fgsm-pgd-cw/https://adversarialml.dev/posts/evasion-attacks-fgsm-pgd-cw/The three foundational gradient-based evasion attacks, what each one actually optimizes, and what the benchmark numbers mean when you're evaluating a defense.Sun, 10 May 2026 00:00:00 GMTevasion-attacksfgsmpgdcarlini-wagneradversarial-examplesadversarial-mlimage-classifiersAdversarial ML Editorialhttps://adversarialml.dev/posts/model-inversion-attacks/https://adversarialml.dev/posts/model-inversion-attacks/From Fredrikson's pharmacogenetics exploit to Geiping's gradient inversion, model inversion attacks recover private training data in ways most MLSun, 10 May 2026 00:00:00 GMTmodel-inversionprivacygradient-inversiontraining-dataadversarial-mlfederated-learningAdversarial ML Editorialhttps://adversarialml.dev/posts/transferability-black-box-attacks/https://adversarialml.dev/posts/transferability-black-box-attacks/Adversarial examples transfer across models with different architectures and training sets. Understanding why changes what you think defenses need toSun, 10 May 2026 00:00:00 GMTtransferabilityblack-box-attacksadversarial-examplesevasionadversarial-mlml-securityAdversarial ML Editorialhttps://adversarialml.dev/posts/certified-robustness-randomized-smoothing/https://adversarialml.dev/posts/certified-robustness-randomized-smoothing/Randomized smoothing gives you a provable robustness radius. Understanding what that certificate means in practice — and where it breaks — is more usefulSat, 09 May 2026 00:00:00 GMTcertified-robustnessrandomized-smoothingadversarial-defenseml-securityformal-verificationAdversarial ML Editorialhttps://adversarialml.dev/posts/training-data-extraction-llms/https://adversarialml.dev/posts/training-data-extraction-llms/Carlini et al. demonstrated verbatim extraction of training data from GPT-2. The results have been widely misread.Fri, 08 May 2026 00:00:00 GMTtraining-data-extractionmemorizationprivacyllm-securitygdprAdversarial ML Editorialhttps://adversarialml.dev/posts/membership-inference-attacks/https://adversarialml.dev/posts/membership-inference-attacks/Shokri et al.'s shadow-model attack is the canonical reference, but the gap between the paper's threat model and a real rate-limited API is wide.Thu, 07 May 2026 00:00:00 GMTmembership-inferenceprivacyml-securityproduction-mlred-teamAdversarial ML Editorialhttps://adversarialml.dev/posts/gcg-class-adversarial-suffix-2026/https://adversarialml.dev/posts/gcg-class-adversarial-suffix-2026/The math, the cost curve, and why optimization-based attacks are now within reach of solo practitioners. With reproducible setup and what defendersThu, 07 May 2026 00:00:00 GMTadversarial-mlgcgoptimization-attacksred-teamalignmentAdversarial ML Editorialhttps://adversarialml.dev/posts/model-extraction-attacks/https://adversarialml.dev/posts/model-extraction-attacks/Query-based model stealing attacks can recover a functionally equivalent model from API access alone. The economics matter more than the technique: here'sThu, 07 May 2026 00:00:00 GMTmodel-extractionmodel-stealingml-securityadversarial-mlapi-securityAdversarial ML Editorial