How much do language models memorize?

We propose a new method for estimating how much a model ``knows'' about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: \textit{unintended memorization}, the information a model contains about a specific dataset, and \textit{generalization}, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point ``grokking'' begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.

Detecting Benchmark Contamination Through Watermarking

Benchmark contamination poses a significant challenge to the reliability of Large Language Models (LLMs) evaluations, as it is difficult to assert whether a model has been trained on a test set. We introduce a solution to this problem by watermarking benchmarks before their release. The embedding involves reformulating the original questions with a watermarked LLM, in a way that does not alter the benchmark utility. During evaluation, we can detect ``radioactivity'', \ie traces that the text watermarks leave in the model during training, using a theoretically grounded statistical test. We test our method by pre-training 1B models from scratch on 10B tokens with controlled benchmark contamination, and validate its effectiveness in detecting contamination on ARC-Easy, ARC-Challenge, and MMLU. Results show similar benchmark utility post-watermarking and successful contamination detection when models are contaminated enough to enhance performance, e.g. $p$-val $=10^{-3}$ for +5$\%$ on ARC-Easy.

Unlocking Visual Secrets: Inverting Features with Diffusion Priors for Image Reconstruction

Inverting visual representations within deep neural networks (DNNs) presents a challenging and important problem in the field of security and privacy for deep learning. The main goal is to invert the features of an unidentified target image generated by a pre-trained DNN, aiming to reconstruct the original image. Feature inversion holds particular significance in understanding the privacy leakage inherent in contemporary split DNN execution techniques, as well as in various applications based on the extracted DNN features. In this paper, we explore the use of diffusion models, a promising technique for image synthesis, to enhance feature inversion quality. We also investigate the potential of incorporating alternative forms of prior knowledge, such as textual prompts and cross-frame temporal correlations, to further improve the quality of inverted features. Our findings reveal that diffusion models can effectively leverage hidden information from the DNN features, resulting in superior reconstruction performance compared to previous methods. This research offers valuable insights into how diffusion models can enhance privacy and security within applications that are reliant on DNN features.

Auditing $f$-Differential Privacy in One Run

Empirical auditing has emerged as a means of catching some of the flaws in the implementation of privacy-preserving algorithms. Existing auditing mechanisms, however, are either computationally inefficient requiring multiple runs of the machine learning algorithms or suboptimal in calculating an empirical privacy. In this work, we present a tight and efficient auditing procedure and analysis that can effectively assess the privacy of mechanisms. Our approach is efficient; similar to the recent work of Steinke, Nasr, and Jagielski (2023), our auditing procedure leverages the randomness of examples in the input dataset and requires only a single run of the target mechanism. And it is more accurate; we provide a novel analysis that enables us to achieve tight empirical privacy estimates by using the hypothesized $f$-DP curve of the mechanism, which provides a more accurate measure of privacy than the traditional $\epsilon,\delta$ differential privacy parameters. We use our auditing procure and analysis to obtain empirical privacy, demonstrating that our auditing procedure delivers tighter privacy estimates.

Aligning LLMs to Be Robust Against Prompt Injection

Large language models (LLMs) are becoming increasingly prevalent in modern software systems, interfacing between the user and the internet to assist with tasks that require advanced language understanding. To accomplish these tasks, the LLM often uses external data sources such as user documents, web retrieval, results from API calls, etc. This opens up new avenues for attackers to manipulate the LLM via prompt injection. Adversarial prompts can be carefully crafted and injected into external data sources to override the user's intended instruction and instead execute a malicious instruction. Prompt injection attacks constitute a major threat to LLM security, making the design and implementation of practical countermeasures of paramount importance. To this end, we show that alignment can be a powerful tool to make LLMs more robust against prompt injection. Our method -- SecAlign -- first builds an alignment dataset by simulating prompt injection attacks and constructing pairs of desirable and undesirable responses. Then, we apply existing alignment techniques to fine-tune the LLM to be robust against these simulated attacks. Our experiments show that SecAlign robustifies the LLM substantially with a negligible hurt on model utility. Moreover, SecAlign's protection generalizes to strong attacks unseen in training. Specifically, the success rate of state-of-the-art GCG-based prompt injections drops from 56% to 2% in Mistral-7B after our alignment process. Our code is released at https://github.com/facebookresearch/SecAlign

Guarantees of confidentiality via Hammersley-Chapman-Robbins bounds

Protecting privacy during inference with deep neural networks is possible by adding noise to the activations in the last layers prior to the final classifiers or other task-specific layers. The activations in such layers are known as "features" (or, less commonly, as "embeddings" or "feature embeddings"). The added noise helps prevent reconstruction of the inputs from the noisy features. Lower bounding the variance of every possible unbiased estimator of the inputs quantifies the confidentiality arising from such added noise. Convenient, computationally tractable bounds are available from classic inequalities of Hammersley and of Chapman and Robbins -- the HCR bounds. Numerical experiments indicate that the HCR bounds are on the precipice of being effectual for small neural nets with the data sets, "MNIST" and "CIFAR-10," which contain 10 classes each for image classification. The HCR bounds appear to be insufficient on their own to guarantee confidentiality of the inputs to inference with standard deep neural nets, "ResNet-18" and "Swin-T," pre-trained on the data set, "ImageNet-1000," which contains 1000 classes. Supplementing the addition of noise to features with other methods for providing confidentiality may be warranted in the case of ImageNet. In all cases, the results reported here limit consideration to amounts of added noise that incur little degradation in the accuracy of classification from the noisy features. Thus, the added noise enhances confidentiality without much reduction in the accuracy on the task of image classification.

Privacy Amplification for the Gaussian Mechanism via Bounded Support

Data-dependent privacy accounting frameworks such as per-instance differential privacy (pDP) and Fisher information loss (FIL) confer fine-grained privacy guarantees for individuals in a fixed training dataset. These guarantees can be desirable compared to vanilla DP in real world settings as they tightly upper-bound the privacy leakage for a $\textit{specific}$ individual in an $\textit{actual}$ dataset, rather than considering worst-case datasets. While these frameworks are beginning to gain popularity, to date, there is a lack of private mechanisms that can fully leverage advantages of data-dependent accounting. To bridge this gap, we propose simple modifications of the Gaussian mechanism with bounded support, showing that they amplify privacy guarantees under data-dependent accounting. Experiments on model training with DP-SGD show that using bounded support Gaussian mechanisms can provide a reduction of the pDP bound $\epsilon$ by as much as 30% without negative effects on model utility.

Private Fine-tuning of Large Language Models with Zeroth-order Optimization

Fine-tuning large pretrained models on private datasets may run the risk of violating privacy. Differential privacy is a framework for mitigating privacy risks by enforcing algorithmic stability. DP-SGD enables training models with private data in a privacy-preserving manner, but raises new obstacles in the form of performance loss and significant engineering challenges. We introduce DP-ZO, a new method for fine-tuning large language models that preserves the privacy of training data by privatizing zeroth-order optimization. A key insight into the design of our method is that the direction of the gradient in SPSA, the zeroth-order algorithm we use, is always random and the only information that depends on private data is the step size, i.e., a scalar. Therefore, we only need to privatize the scalar step size, which is memory-efficient. DP-ZO, which can be instantiated with either Laplace or Gaussian noise, provides a strong privacy-utility trade-off across different tasks, and model sizes, under conservative privacy budgets. One noteworthy result is that DP-ZO exhibits just $1.86\%$ performance degradation due to privacy at $(1,10^{-5})$-DP when fine-tuning OPT-66B on 1000 training samples from SQuAD.

Publicly Detectable Watermarking for Language Models

We construct the first provable watermarking scheme for language models with public detectability or verifiability: we use a private key for watermarking and a public key for watermark detection. Our protocol is the first watermarking scheme that does not embed a statistical signal in generated text. Rather, we directly embed a publicly-verifiable cryptographic signature using a form of rejection sampling. We show that our construction meets strong formal security guarantees and preserves many desirable properties found in schemes in the private-key watermarking setting. In particular, our watermarking scheme retains distortion-freeness and model agnosticity. We implement our scheme and make empirical measurements over open models in the 7B parameter range. Our experiments suggest that our watermarking scheme meets our formal claims while preserving text quality.

A Randomized Approach for Tight Privacy Accounting

Bounding privacy leakage over compositions, i.e., privacy accounting, is a key challenge in differential privacy (DP). However, the privacy parameter ($\varepsilon$ or $\delta$) is often easy to estimate but hard to bound. In this paper, we propose a new differential privacy paradigm called estimate-verify-release (EVR), which addresses the challenges of providing a strict upper bound for privacy parameter in DP compositions by converting an estimate of privacy parameter into a formal guarantee. The EVR paradigm first estimates the privacy parameter of a mechanism, then verifies whether it meets this guarantee, and finally releases the query output based on the verification result. The core component of the EVR is privacy verification. We develop a randomized privacy verifier using Monte Carlo (MC) technique. Furthermore, we propose an MC-based DP accountant that outperforms existing DP accounting techniques in terms of accuracy and efficiency. Our empirical evaluation shows the newly proposed EVR paradigm improves the utility-privacy tradeoff for privacy-preserving machine learning.