Christopher A. Choquette-Choo

Technical Reports and Production Deployments

We show that production aligned models can be provide a false-sense of privacy.We show a vulnerability in Chat-GPT that lets us extract a lot of data. We scalably test for extractable memorization in many open-source models.

We prove a generic privacy amplification technique for DP-FTRL.

We prove bounds on DP-FTRL that show when it is better than DP-SGD. We design new algorithms that are more efficient and as performant.

We show new attacks that can extract sensitive PII from LLMs with high success and minimal assumptions.

We show new attacks against user's for LLMs tuned on sensitive user data that achieve high success. We call these user inference attacks.

We show that an entire system---not just an ML model---need to be considered to ensure privacy. We introduce new attacks against a variety of systems components that can invalidate guarantees on just an ML model.

We discuss and provide insights/opinions on the intersection of generative AI and law.

We introduce a new multilingual dataset, improve multilingual models, and show that translation tasks can be vulnerable to new privacy attacks.

We achieve new state-of-the-art privacy-utility tradeoffs in DP ML. We outperform DP-SGD without any need for privacy amplification.

We show how to achieve impressive compression rates in the online setting, i.e., without prior tuning. Our best algorithms can achieve close to the best-known compression rates.

We show that DP-FTRL is strictly better than DP-SGD. In doing so, we get new state-of-the-art.

We introduce attacks against multi-modal models that subvert alignment techniques. In doing so, we show that current NLP attacks are not sufficiently powerful to evaluate adversarial alignment.

We show that modern large-scale datasets can be cost-effectively poisoned.

We show that, without formal guarantees, distillation provides limited privacy benefits. We also show the first attack where a model can leak information about private data without ever having been queried on that private data.

We show that models can paraphrase memorizations. Thus, models can evade filters designed specifically to prevent verbatim memorization, like those implemented for CoPilot.

We break current Proof-of-Learning schemes. Though we design schemes that are robust to all current attacks, we show that it is an open problem to provide formal guarantees.

Do proper hyper-parameter tuning with DP. Check how many layers you should tune.

We characterize the fundamental communication costs of Federated Learning (FL) under Secure Aggregation (SecAgg) and Differential Privacy (DP), two privacy-preserving mechanisms that are commonly used with FL. We prove our optimality for worst-case settings which provides significant improvements over prior work, and show that improvements can be made under additional assumptions, e.g., data sparisty. Extensive empirical evaluations support our claims, showing costs as low as 1.2 bits per parameter on Stack Overflow with <4% relative decrease in test-time model accuracy.

We show that in the worst-case, differentially-private federated learning with secure aggregation requires Ω(d) bits. Despite this, we discuss how to leverage near-sparsity to compress updates by more than 50x with modest noise multipliers of 0.4 by using sketching.

How can we prove that a machine learning model owner trained their model? We define the problem of Proof-of-Learning (PoL) in machine learning and provide a method for it that is robust to several spoofing attacks. This protocol enables model ownership verification and robustness against byzantines workers (in a distributed learning setting).

We design a protocol for collaborative learning that ensures both the privacy of the training data and confidentiality of the test data in a collaborative setup. Our protocol can provide several percentage-points improvement to models, especially on subpopulations that the model underperforms on, with a modest privacy budget usage of less than 20. Unlike prior work, we enable collaborative learning of heterogeneous models amongst participants. Unlike differentially private federated learning, which requires ~1 million participants, our protocol can work in regimes of ~100 participants.

What defenses properly defend against all membership inference threats? We expose and show that confidence-masking -- defensive obfuscation of confidence-vectors -- is not a viable defense to Membership Inference. We do this by introducing (3) label-only attacks, which bypass this defense and match typical confidence-vector attacks. In an extensive evaluation of defenses,including the first evaluation of data augmentations and transfer learning as defenses, we further show that Differential Privacy can defend against average- and worse-case Membership Inference attacks.

How can we enable an IP owner to reliably claim ownership of a stolen model? We explore entangling of watermarks to task data to ensure that stolen models learn these watermarks as well. Our improved watermarks enable IP owners to claim ownership with 95% confidence in less than 10 queries to the stolen model.

How can we enable efficient and guaranteed retraining of machine learning models? We define requirements for machine unlearning and study a stricter unlearning requirement whereby unlearning a datapoint is guaranteed to be equivalent to as if we had never trained on it. To this end, we improve on the naive retraining-from-scratch approach to provide a better accuracy-efficieny tradeoff. We also study how a priori knowledge of the distribution of requests can further improve efficiency.

How can we triage bugs more effectively? By utilizing a model's own knowledge of an analogous lower-dimensionality solution-space (predicting the correct team assignment), we can achieve higher accuracies on the higher-dimensionality solution-space (predicting the correct engineer assignment).