Dimitar I. Dimitrov

Dimitar I. Dimitrov, Mislav Balunović, Nikola Konstantinov et al.

Recent attacks have shown that user data can be recovered from FedSGD updates, thus breaking privacy. However, these attacks are of limited practical relevance as federated learning typically uses the FedAvg algorithm. Compared to FedSGD, recovering data from FedAvg updates is much harder as: (i) the updates are computed at unobserved intermediate network weights, (ii) a large number of batches are used, and (iii) labels and network weights vary simultaneously across client steps. In this work, we propose a new optimization-based attack which successfully attacks FedAvg by addressing the above challenges. First, we solve the optimization problem using automatic differentiation that forces a simulation of the client's update that generates the unobserved parameters for the recovered labels and inputs to match the received client update. Second, we address the large number of batches by relating images from different epochs with a permutation invariant prior. Third, we recover the labels by estimating the parameters of existing FedSGD attacks at every FedAvg step. On the popular FEMNIST dataset, we demonstrate that on average we successfully recover >45% of the client's images from realistic FedAvg updates computed on 10 local epochs of 10 batches each with 5 images, compared to only <10% using the baseline. Our findings show many real-world federated learning implementations based on FedAvg are vulnerable.

Nikola Jovanović, Mislav Balunović, Dimitar I. Dimitrov et al.

Fair representation learning (FRL) is a popular class of methods aiming to produce fair classifiers via data preprocessing. Recent regulatory directives stress the need for FRL methods that provide practical certificates, i.e., provable upper bounds on the unfairness of any downstream classifier trained on preprocessed data, which directly provides assurance in a practical scenario. Creating such FRL methods is an important challenge that remains unsolved. In this work, we address that challenge and introduce FARE (Fairness with Restricted Encoders), the first FRL method with practical fairness certificates. FARE is based on our key insight that restricting the representation space of the encoder enables the derivation of practical guarantees, while still permitting favorable accuracy-fairness tradeoffs for suitable instantiations, such as one we propose based on fair trees. To produce a practical certificate, we develop and apply a statistical procedure that computes a finite sample high-confidence upper bound on the unfairness of any downstream classifier trained on FARE embeddings. In our comprehensive experimental evaluation, we demonstrate that FARE produces practical certificates that are tight and often even comparable with purely empirical results obtained by prior methods, which establishes the practical value of our approach.

Mark Vero, Mislav Balunović, Dimitar I. Dimitrov et al.

While federated learning (FL) promises to preserve privacy, recent works in the image and text domains have shown that training updates leak private client data. However, most high-stakes applications of FL (e.g., in healthcare and finance) use tabular data, where the risk of data leakage has not yet been explored. A successful attack for tabular data must address two key challenges unique to the domain: (i) obtaining a solution to a high-variance mixed discrete-continuous optimization problem, and (ii) enabling human assessment of the reconstruction as unlike for image and text data, direct human inspection is not possible. In this work we address these challenges and propose TabLeak, the first comprehensive reconstruction attack on tabular data. TabLeak is based on two key contributions: (i) a method which leverages a softmax relaxation and pooled ensembling to solve the optimization problem, and (ii) an entropy-based uncertainty quantification scheme to enable human assessment. We evaluate TabLeak on four tabular datasets for both FedSGD and FedAvg training protocols, and show that it successfully breaks several settings previously deemed safe. For instance, we extract large subsets of private data at >90% accuracy even at the large batch size of 128. Our findings demonstrate that current high-stakes tabular FL is excessively vulnerable to leakage attacks.

Kostadin Garov, Dimitar I. Dimitrov, Nikola Jovanović et al.

Malicious server (MS) attacks have enabled the scaling of data stealing in federated learning to large batch sizes and secure aggregation, settings previously considered private. However, many concerns regarding the client-side detectability of MS attacks were raised, questioning their practicality. In this work, for the first time, we thoroughly study client-side detectability. We first demonstrate that all prior MS attacks are detectable by principled checks, and formulate a necessary set of requirements that a practical MS attack must satisfy. Next, we propose SEER, a novel attack framework that satisfies these requirements. The key insight of SEER is the use of a secret decoder, jointly trained with the shared model. We show that SEER can steal user data from gradients of realistic networks, even for large batch sizes of up to 512 and under secure aggregation. Our work is a promising step towards assessing the true vulnerability of federated learning in real-world settings.

Csaba Dékány, Stefan Balauca, Robin Staab et al.

Despite recent efforts in Large Language Model (LLM) safety and alignment, current adversarial attacks on frontier LLMs can still consistently force harmful generations. Although adversarial training has been widely studied and shown to significantly improve the robustness of traditional machine learning models, its strengths and weaknesses in the context of LLMs are less understood. Specifically, while existing discrete adversarial attacks are effective at producing harmful content, training LLMs with concrete adversarial prompts is often computationally expensive, leading to reliance on continuous relaxations. At the same time, despite their effectiveness and generalization capabilities, training with continuous perturbations does not always capture the full spectrum of vulnerabilities exploited by discrete attacks. In this work, we aim to bridge this gap by introducing MixAT, a novel method that combines stronger discrete and faster continuous attacks during training. We rigorously evaluate MixAT across a wide spectrum of state-of-the-art attacks, proposing the At Least One Attack Success Rate (ALO-ASR) metric to capture the worst-case vulnerability of models. We show MixAT achieves substantially better robustness (ALO-ASR < 20%) compared to prior defenses (ALO-ASR > 50%), while maintaining a runtime comparable to methods based on continuous relaxations. We further analyze MixAT in realistic deployment settings, exploring how chat templates, quantization, low-rank adapters, and temperature affect both adversarial training and evaluation, revealing additional blind spots in current methodologies. Our results demonstrate that MixAT's discrete-continuous defense offers a principled and superior robustness-accuracy tradeoff with minimal computational overhead, highlighting its promise for building safer LLMs. We provide our code and models at https://github.com/insait-institute/MixAT.

Ivo Petrov, Jasper Dekoninck, Dimitar I. Dimitrov et al.

Large language models (LLMs) have become capable mathematical problem-solvers, often producing correct proofs for challenging problems. However, correctness alone is not sufficient: mathematical proofs should also be clear, concise, insightful, and transferable to other problems. While this proof quality is subjective and depends on the reader and context, many of its components are concrete and broadly valued. In this work, we identify such components and introduce ProofRank, a benchmark curated from challenging mathematical competitions. ProofRank evaluates several scalable proxies of proof quality: (i) conciseness, measuring whether proofs avoid unnecessary steps; (ii) computational ease, measuring the extent to which a proof relies on tedious calculations; (iii) cognitive simplicity, measuring how accessible the used proof techniques are; (iv) diversity, measuring how varied a model's proofs for a single problem are; and (v) adaptivity, measuring whether a model can follow a specified proof technique. Across models, we find substantial differences in proof quality that are not captured by correctness-only benchmarks. We also observe significant trade-offs between proof-quality metrics and correctness, suggesting that future evaluations of mathematical reasoning should measure how useful LLM-generated proofs are.

Marc Fischer, Christian Sprecher, Dimitar I. Dimitrov et al.

Existing neural network verifiers compute a proof that each input is handled correctly under a given perturbation by propagating a symbolic abstraction of reachable values at each layer. This process is repeated from scratch independently for each input (e.g., image) and perturbation (e.g., rotation), leading to an expensive overall proof effort when handling an entire dataset. In this work, we introduce a new method for reducing this verification cost without losing precision based on a key insight that abstractions obtained at intermediate layers for different inputs and perturbations can overlap or contain each other. Leveraging our insight, we introduce the general concept of shared certificates, enabling proof effort reuse across multiple inputs to reduce overall verification costs. We perform an extensive experimental evaluation to demonstrate the effectiveness of shared certificates in reducing the verification cost on a range of datasets and attack specifications on image classifiers including the popular patch and geometric perturbations. We release our implementation at https://github.com/eth-sri/proof-sharing.

Dimitar I. Dimitrov, Maximilian Baader, Mark Niklas Müller et al.

Federated learning is a framework for collaborative machine learning where clients only share gradient updates and not their private data with a server. However, it was recently shown that gradient inversion attacks can reconstruct this data from the shared gradients. In the important honest-but-curious setting, existing attacks enable exact reconstruction only for batch size of $b=1$, with larger batches permitting only approximate reconstruction. In this work, we propose SPEAR, the first algorithm reconstructing whole batches with $b >1$ exactly. SPEAR combines insights into the explicit low-rank structure of gradients with a sampling-based algorithm. Crucially, we leverage ReLU-induced gradient sparsity to precisely filter out large numbers of incorrect samples, making a final reconstruction step tractable. We provide an efficient GPU implementation for fully connected networks and show that it recovers high-dimensional ImageNet inputs in batches of up to $b \lesssim 25$ exactly while scaling to large networks. Finally, we show theoretically that much larger batches can be reconstructed with high probability given exponential time.

Ivo Petrov, Dimitar I. Dimitrov, Maximilian Baader et al. · eth-zurich

Federated learning works by aggregating locally computed gradients from multiple clients, thus enabling collaborative training without sharing private client data. However, prior work has shown that the data can actually be recovered by the server using so-called gradient inversion attacks. While these attacks perform well when applied on images, they are limited in the text domain and only permit approximate reconstruction of small batches and short input sequences. In this work, we propose DAGER, the first algorithm to recover whole batches of input text exactly. DAGER leverages the low-rank structure of self-attention layer gradients and the discrete nature of token embeddings to efficiently check if a given token sequence is part of the client data. We use this check to exactly recover full batches in the honest-but-curious setting without any prior on the data for both encoder- and decoder-based architectures using exhaustive heuristic search and a greedy approach, respectively. We provide an efficient GPU implementation of DAGER and show experimentally that it recovers full batches of size up to 128 on large language models (LLMs), beating prior attacks in speed (20x at same batch size), scalability (10x larger batches), and reconstruction quality (ROUGE-1/2 > 0.99).

Maria Drencheva, Ivo Petrov, Maximilian Baader et al.

Federated learning claims to enable collaborative model training among multiple clients with data privacy by transmitting gradient updates instead of the actual client data. However, recent studies have shown the client privacy is still at risk due to the, so called, gradient inversion attacks which can precisely reconstruct clients' text and image data from the shared gradient updates. While these attacks demonstrate severe privacy risks for certain domains and architectures, the vulnerability of other commonly-used data types, such as graph-structured data, remain under-explored. To bridge this gap, we present GRAIN, the first exact gradient inversion attack on graph data in the honest-but-curious setting that recovers both the structure of the graph and the associated node features. Concretely, we focus on Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) -- two of the most widely used frameworks for learning on graphs. Our method first utilizes the low-rank structure of GNN gradients to efficiently reconstruct and filter the client subgraphs which are then joined to complete the input graph. We evaluate our approach on molecular, citation, and social network datasets using our novel metric. We show that GRAIN reconstructs up to 80% of all graphs exactly, significantly outperforming the baseline, which achieves up to 20% correctly positioned nodes.

Anton Alexandrov, Veselin Raychev, Dimitar I. Dimitrov et al.

We present BgGPT-Gemma-2-27B-Instruct and BgGPT-Gemma-2-9B-Instruct: continually pretrained and fine-tuned versions of Google's Gemma-2 models, specifically optimized for Bulgarian language understanding and generation. Leveraging Gemma-2's multilingual capabilities and over 100 billion tokens of Bulgarian and English text data, our models demonstrate strong performance in Bulgarian language tasks, setting a new standard for language-specific AI models. Our approach maintains the robust capabilities of the original Gemma-2 models, ensuring that the English language performance remains intact. To preserve the base model capabilities, we incorporate continual learning strategies based on recent Branch-and-Merge techniques as well as thorough curation and selection of training data. We provide detailed insights into our methodology, including the release of model weights with a commercial-friendly license, enabling broader adoption by researchers, companies, and hobbyists. Further, we establish a comprehensive set of benchmarks based on non-public educational data sources to evaluate models on Bulgarian language tasks as well as safety and chat capabilities. Our findings demonstrate the effectiveness of fine-tuning state-of-the-art models like Gemma 2 to enhance language-specific AI applications while maintaining cross-lingual capabilities.

Alexander Bakarsky, Dimitar I. Dimitrov, Maximilian Baader et al.

Federated Learning has seen an increased deployment in real-world scenarios recently, as it enables the distributed training of machine learning models without explicit data sharing between individual clients. Yet, the introduction of the so-called gradient inversion attacks has fundamentally challenged its privacy-preserving properties. Unfortunately, as these attacks mostly rely on direct data optimization without any formal guarantees, the vulnerability of real-world systems remains in dispute and requires tedious testing for each new federated deployment. To overcome these issues, recently the SPEAR attack was introduced, which is based on a theoretical analysis of the gradients of linear layers with ReLU activations. While SPEAR is an important theoretical breakthrough, the attack's practicality was severely limited by its exponential runtime in the batch size b. In this work, we fill this gap by applying State-of-the-Art techniques from Sparsely-Used Dictionary Learning to make the problem of gradient inversion on linear layers with ReLU activations tractable. Our experiments demonstrate that our new attack, SPEAR++, retains all desirable properties of SPEAR, such as robustness to DP noise and FedAvg aggregation, while being applicable to 10x bigger batch sizes.

Mislav Balunović, Dimitar I. Dimitrov, Nikola Jovanović et al.

Recent work shows that sensitive user data can be reconstructed from gradient updates, breaking the key privacy promise of federated learning. While success was demonstrated primarily on image data, these methods do not directly transfer to other domains such as text. In this work, we propose LAMP, a novel attack tailored to textual data, that successfully reconstructs original text from gradients. Our attack is based on two key insights: (i) modeling prior text probability with an auxiliary language model, guiding the search towards more natural text, and (ii) alternating continuous and discrete optimization, which minimizes reconstruction loss on embeddings, while avoiding local minima by applying discrete text transformations. Our experiments demonstrate that LAMP is significantly more effective than prior work: it reconstructs 5x more bigrams and 23% longer subsequences on average. Moreover, we are the first to recover inputs from batch sizes larger than 1 for textual models. These findings indicate that gradient updates of models operating on textual data leak more information than previously thought.

Mislav Balunović, Dimitar I. Dimitrov, Robin Staab et al.

Federated learning is an established method for training machine learning models without sharing training data. However, recent work has shown that it cannot guarantee data privacy as shared gradients can still leak sensitive information. To formalize the problem of gradient leakage, we propose a theoretical framework that enables, for the first time, analysis of the Bayes optimal adversary phrased as an optimization problem. We demonstrate that existing leakage attacks can be seen as approximations of this optimal adversary with different assumptions on the probability distributions of the input data and gradients. Our experiments confirm the effectiveness of the Bayes optimal adversary when it has knowledge of the underlying distribution. Further, our experimental evaluation shows that several existing heuristic defenses are not effective against stronger attacks, especially early in the training process. Thus, our findings indicate that the construction of more effective defenses and their evaluation remains an open problem.

Dimitar I. Dimitrov, Gagandeep Singh, Timon Gehr et al.

We introduce the concept of provably robust adversarial examples for deep neural networks - connected input regions constructed from standard adversarial examples which are guaranteed to be robust to a set of real-world perturbations (such as changes in pixel intensity and geometric transformations). We present a novel method called PARADE for generating these regions in a scalable manner which works by iteratively refining the region initially obtained via sampling until a refined region is certified to be adversarial with existing state-of-the-art verifiers. At each step, a novel optimization procedure is applied to maximize the region's volume under the constraint that the convex relaxation of the network behavior with respect to the region implies a chosen bound on the certification objective. Our experimental evaluation shows the effectiveness of PARADE: it successfully finds large provably robust regions including ones containing $\approx 10^{573}$ adversarial examples for pixel intensity and $\approx 10^{599}$ for geometric perturbations. The provability enables our robust examples to be significantly more effective against state-of-the-art defenses based on randomized smoothing than the individual attacks used to construct the regions.