Andreas Terzis

Nicholas Carlini, Matthew Jagielski, Christopher A. Choquette-Choo et al. · deepmind, eth-zurich

Deep learning models are often trained on distributed, web-scale datasets crawled from the internet. In this paper, we introduce two new dataset poisoning attacks that intentionally introduce malicious examples to a model's performance. Our attacks are immediately practical and could, today, poison 10 popular datasets. Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator's initial view of the dataset differs from the view downloaded by subsequent clients. By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD. Our second attack, frontrunning poisoning, targets web-scale datasets that periodically snapshot crowd-sourced content -- such as Wikipedia -- where an attacker only needs a time-limited window to inject malicious examples. In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.

Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang et al. · eth-zurich

Machine learning models trained on private datasets have been shown to leak their private data. While recent work has found that the average data point is rarely leaked, the outlier samples are frequently subject to memorization and, consequently, privacy leakage. We demonstrate and analyse an Onion Effect of memorization: removing the "layer" of outlier points that are most vulnerable to a privacy attack exposes a new layer of previously-safe points to the same attack. We perform several experiments to study this effect, and understand why it occurs. The existence of this effect has various consequences. For example, it suggests that proposals to defend against memorization without training with rigorous privacy guarantees are unlikely to be effective. Further, it suggests that privacy-enhancing technologies such as machine unlearning could actually harm the privacy of other users.

Milad Nasr, Jamie Hayes, Thomas Steinke et al. · eth-zurich

Auditing mechanisms for differential privacy use probabilistic means to empirically estimate the privacy level of an algorithm. For private machine learning, existing auditing mechanisms are tight: the empirical privacy estimate (nearly) matches the algorithm's provable privacy guarantee. But these auditing techniques suffer from two limitations. First, they only give tight estimates under implausible worst-case assumptions (e.g., a fully adversarial dataset). Second, they require thousands or millions of training runs to produce non-trivial statistical estimates of the privacy leakage. This work addresses both issues. We design an improved auditing scheme that yields tight privacy estimates for natural (not adversarially crafted) datasets -- if the adversary can see all model updates during training. Prior auditing works rely on the same assumption, which is permitted under the standard differential privacy threat model. This threat model is also applicable, e.g., in federated learning settings. Moreover, our auditing scheme requires only two training runs (instead of thousands) to produce tight privacy estimates, by adapting recent advances in tight composition theorems for differential privacy. We demonstrate the utility of our improved auditing schemes by surfacing implementation bugs in private machine learning code that eluded prior auditing techniques.

Alexey Kurakin, Natalia Ponomareva, Umar Syed et al.

Differentially private training algorithms like DP-SGD protect sensitive training data by ensuring that trained models do not reveal private information. An alternative approach, which this paper studies, is to use a sensitive dataset to generate synthetic data that is differentially private with respect to the original data, and then non-privately training a model on the synthetic data. Doing so has several advantages: synthetic data can be reused for other tasks (including for hyper parameter tuning), retained indefinitely, and shared with third parties without sacrificing privacy. However, generating private synthetic data is much harder than training a private model. To improve performance on text data, recent work has utilized public data by starting with a pre-trained generative language model and privately fine-tuning it on sensitive data. This model can be used to sample a DP synthetic dataset. While this strategy seems straightforward, executing it has proven problematic. Previous approaches either show significant performance loss, or have, as we show, critical design flaws. In this paper we demonstrate that a proper training objective along with tuning fewer parameters results in excellent DP synthetic data quality. Our approach is competitive with direct DP-training of downstream classifiers in terms of performance on downstream tasks. Further, we demonstrate that our DP synthetic data is not only useful for downstream classifier training, but also to tune those same models.

Kareem Amin, Alex Bie, Weiwei Kong et al.

We present an approach for generating differentially private synthetic text using large language models (LLMs), via private prediction. In the private prediction framework, we only require the output synthetic data to satisfy differential privacy guarantees. This is in contrast to approaches that train a generative model on potentially sensitive user-supplied source data and seek to ensure the model itself is safe to release. We prompt a pretrained LLM with source data, but ensure that next-token predictions are made with differential privacy guarantees. Previous work in this paradigm reported generating a small number of examples (<10) at reasonable privacy levels, an amount of data that is useful only for downstream in-context learning or prompting. In contrast, we make changes that allow us to generate thousands of high-quality synthetic data points, greatly expanding the set of potential applications. Our improvements come from an improved privacy analysis and a better private selection mechanism, which makes use of the equivalence between the softmax layer for sampling tokens in LLMs and the exponential mechanism. Furthermore, we introduce a novel use of public predictions via the sparse vector technique, in which we do not pay privacy costs for tokens that are predictable without sensitive data; we find this to be particularly effective for structured data.

Natalia Ponomareva, Zheng Xu, H. Brendan McMahan et al.

High quality data is needed to unlock the full potential of AI for end users. However finding new sources of such data is getting harder: most publicly-available human generated data will soon have been used. Additionally, publicly available data often is not representative of users of a particular system -- for example, a research speech dataset of contractors interacting with an AI assistant will likely be more homogeneous, well articulated and self-censored than real world commands that end users will issue. Therefore unlocking high-quality data grounded in real user interactions is of vital interest. However, the direct use of user data comes with significant privacy risks. Differential Privacy (DP) is a well established framework for reasoning about and limiting information leakage, and is a gold standard for protecting user privacy. The focus of this work, \emph{Differentially Private Synthetic data}, refers to synthetic data that preserves the overall trends of source data,, while providing strong privacy guarantees to individuals that contributed to the source dataset. DP synthetic data can unlock the value of datasets that have previously been inaccessible due to privacy concerns and can replace the use of sensitive datasets that previously have only had rudimentary protections like ad-hoc rule-based anonymization. In this paper we explore the full suite of techniques surrounding DP synthetic data, the types of privacy protections they offer and the state-of-the-art for various modalities (image, tabular, text and decentralized). We outline all the components needed in a system that generates DP synthetic data, from sensitive data handling and preparation, to tracking the use and empirical privacy testing. We hope that work will result in increased adoption of DP synthetic data, spur additional research and increase trust in DP synthetic data approaches.

Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan et al. · deepmind, eth-zurich

Large Language Models (LLMs) are increasingly deployed in agentic systems that interact with an untrusted environment. However, LLM agents are vulnerable to prompt injection attacks when handling untrusted data. In this paper we propose CaMeL, a robust defense that creates a protective system layer around the LLM, securing it even when underlying models are susceptible to attacks. To operate, CaMeL explicitly extracts the control and data flows from the (trusted) query; therefore, the untrusted data retrieved by the LLM can never impact the program flow. To further improve security, CaMeL uses a notion of a capability to prevent the exfiltration of private data over unauthorized data flows by enforcing security policies when tools are called. We demonstrate effectiveness of CaMeL by solving $77\%$ of tasks with provable security (compared to $84\%$ with an undefended system) in AgentDojo. We release CaMeL at https://github.com/google-research/camel-prompt-injection.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Milad Nasr, Yanick Fratantonio, Luca Invernizzi et al.

As deep learning models become widely deployed as components within larger production systems, their individual shortcomings can create system-level vulnerabilities with real-world impact. This paper studies how adversarial attacks targeting an ML component can degrade or bypass an entire production-grade malware detection system, performing a case study analysis of Gmail's pipeline where file-type identification relies on a ML model. The malware detection pipeline in use by Gmail contains a machine learning model that routes each potential malware sample to a specialized malware classifier to improve accuracy and performance. This model, called Magika, has been open sourced. By designing adversarial examples that fool Magika, we can cause the production malware service to incorrectly route malware to an unsuitable malware detector thereby increasing our chance of evading detection. Specifically, by changing just 13 bytes of a malware sample, we can successfully evade Magika in 90% of cases and thereby allow us to send malware files over Gmail. We then turn our attention to defenses, and develop an approach to mitigate the severity of these types of attacks. For our defended production model, a highly resourced adversary requires 50 bytes to achieve just a 20% attack success rate. We implement this defense, and, thanks to a collaboration with Google engineers, it has already been deployed in production for the Gmail classifier.

Florian Tramer, Andreas Terzis, Thomas Steinke et al.

Differential Privacy can provide provable privacy guarantees for training data in machine learning. However, the presence of proofs does not preclude the presence of errors. Inspired by recent advances in auditing which have been used for estimating lower bounds on differentially private algorithms, here we show that auditing can also be used to find flaws in (purportedly) differentially private schemes. In this case study, we audit a recent open source implementation of a differentially private deep learning algorithm and find, with 99.99999999% confidence, that the implementation does not satisfy the claimed differential privacy guarantee.

Alexey Kurakin, Shuang Song, Steve Chien et al.

Differential privacy (DP) is the de facto standard for training machine learning (ML) models, including neural networks, while ensuring the privacy of individual examples in the training set. Despite a rich literature on how to train ML models with differential privacy, it remains extremely challenging to train real-life, large neural networks with both reasonable accuracy and privacy. We set out to investigate how to do this, using ImageNet image classification as a poster example of an ML task that is very challenging to resolve accurately with DP right now. This paper shares initial lessons from our effort, in the hope that it will inspire and inform other researchers to explore DP training at scale. We show approaches that help make DP training faster, as well as model types and settings of the training process that tend to work better in the DP setting. Combined, the methods we discuss let us train a Resnet-18 with DP to $47.9\%$ accuracy and privacy parameters $ε= 10, δ= 10^{-6}$. This is a significant improvement over "naive" DP training of ImageNet models, but a far cry from the $75\%$ accuracy that can be obtained by the same network without privacy. The model we use was pretrained on the Places365 data set as a starting point. We share our code at https://github.com/google-research/dp-imagenet, calling for others to build upon this new baseline to further improve DP at scale.

Chongyang Shi, Sharon Lin, Shuang Song et al. · deepmind

Gemini is increasingly used to perform tasks on behalf of users, where function-calling and tool-use capabilities enable the model to access user data. Some tools, however, require access to untrusted data introducing risk. Adversaries can embed malicious instructions in untrusted data which cause the model to deviate from the user's expectations and mishandle their data or permissions. In this report, we set out Google DeepMind's approach to evaluating the adversarial robustness of Gemini models and describe the main lessons learned from the process. We test how Gemini performs against a sophisticated adversary through an adversarial evaluation framework, which deploys a suite of adaptive attack techniques to run continuously against past, current, and future versions of Gemini. We describe how these ongoing evaluations directly help make Gemini more resilient against manipulation.

Milad Nasr, Nicholas Carlini, Chawin Sitawarin et al. · eth-zurich

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed. Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques-gradient descent, reinforcement learning, random search, and human-guided exploration-we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

A. Feder Cooper, Christopher A. Choquette-Choo, Miranda Bogen et al. · deepmind

"Machine unlearning" is a popular proposed solution for mitigating the existence of content in an AI model that is problematic for legal or moral reasons, including privacy, copyright, safety, and more. For example, unlearning is often invoked as a solution for removing the effects of specific information from a generative-AI model's parameters, e.g., a particular individual's personal data or the inclusion of copyrighted content in the model's training data. Unlearning is also proposed as a way to prevent a model from generating targeted types of information in its outputs, e.g., generations that closely resemble a particular individual's data or reflect the concept of "Spiderman." Both of these goals--the targeted removal of information from a model and the targeted suppression of information from a model's outputs--present various technical and substantive challenges. We provide a framework for ML researchers and policymakers to think rigorously about these challenges, identifying several mismatches between the goals of unlearning and feasible implementations. These mismatches explain why unlearning is not a general-purpose solution for circumscribing generative-AI model behavior in service of broader positive impact.

Nicholas Carlini, Steve Chien, Milad Nasr et al.

A membership inference attack allows an adversary to query a trained machine learning model to predict whether or not a particular example was contained in the model's training dataset. These attacks are currently evaluated using average-case "accuracy" metrics that fail to characterize whether the attack can confidently identify any members of the training set. We argue that attacks should instead be evaluated by computing their true-positive rate at low (e.g., <0.1%) false-positive rates, and find most prior attacks perform poorly when evaluated in this way. To address this we develop a Likelihood Ratio Attack (LiRA) that carefully combines multiple ideas from the literature. Our attack is 10x more powerful at low false-positive rates, and also strictly dominates prior attacks on existing metrics.

Nicholas Carlini, Andreas Terzis

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.