←
Mohammad Yaghini14 posts
Mohammad Yaghini avatar

Mohammad Yaghini

@MYaghini
◷ Joined July 2011▧ 2 Photos and videos
Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Jul 9

Check out the paper for more details and ablations! Joint work with @AerniMichael, Junrui (Arielle) Zhang, @NicolasPapernot, @florian_tramer See you today at 2:30PM at our poster at #ICML 2026! (Hall A) 🌐 Website: cleverhans.io/optifluence/ 💻 Code: github.com/cleverhans-lab/op…

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Jul 9

How does it work? 1. Select a high-influence training sample as the initial canary. 2. Directly optimize the canary's detectability over the entire training procedure (memory-efficient!) OptiFluence is thus independent of architecture or mechanism.

Post attachment
Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Jul 9

Transferability is crucial for optimized canaries. It allows auditors to create canaries cheaply on a small model to evaluate much larger architectures. Regulators can hence use OptiFluence for privacy audits without access to proprietary training algorithms.

Post attachment
Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Jul 9

⚙️ Most privacy auditing focuses on the mechanism. 🖼️ OptiFluence optimizes the canary instead! Our scalable framework produces canaries with 🔎 near-perfect detection rates (99% TPR@0.1% FPR) 🤸 transferability between architectures This enables practical regulatory audits!

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

The privacy-fairness tension is considerably weaker than previously believed. It appears largely to be an artifact of algorithm design. 🌐 cleverhans.io/dp-raco 📄 iclr.cc/virtual/2026/poster/… 💻 github.com/cleverhans-lab/dp…

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

Unlike regularization-based methods, RaCO-DP enforces bounds directly by construction: no hyperparameter tuning, no extra privacy cost. Also scales to many simultaneous constraints without degradation, and runs three orders of magnitude faster than the prior SOTA, DP-FERMI.

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

On tabular data (Adult, Credit-Card, Parkinsons, ACSEmployment), RaCO-DP closes the gap to non-private models on the accuracy-fairness Pareto frontier. On ResNet16 + CelebA: 90% accuracy, 10% demographic disparity at ε = 1. To the best of our knowledge, not previously achievable.

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

But rate constraints have structure: they decompose into a partinition, bounding each sample to a fixed number of terms. RaCO-DP: privatize a prediction histogram once per step. Gradients and constraint values follow via post-processing, for no extra cost over the standard DP-SGD

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

Standard DP-SGD decomposes cleanly per example: one gradient term per person, clippable and privatizable. Naively adding any regularizer to enforce a constraint gives each sample up to |D|+1 gradient terms, requiring noise proportional to dataset size and destroying utility.

Mohammad Yaghini avatar
Mohammad Yaghini@MYaghini·Apr 25

Can ML models be trained under fairness constraints with formal DP guarantees, without sacrificing utility? New paper at ICLR 2026: RaCO-DP Poster: Pavilion 4 (later today!) 🧵 (w/ @tudorcebere, Michael Menart, Aurélien Bellet, @NicolasPapernot )