François Taı̈ani

Timothée Chauvin, Erwan Le Merrer, François Taïani et al.

When using an LLM through an API provider, users expect the served model to remain consistent over time, a property crucial for the reliability of downstream applications and the reproducibility of research. Existing audit methods are too costly to apply at regular time intervals to the wide range of available LLM APIs. This means that model updates are left largely unmonitored in practice. In this work, we show that while LLM log probabilities (logprobs) are usually non-deterministic, they can still be used as the basis for cost-effective continuous monitoring of LLM APIs. We apply a simple statistical test based on the average value of each token logprob, requesting only a single token of output. This is enough to detect changes as small as one step of fine-tuning, making this approach more sensitive than existing methods while being 1,000x cheaper. We introduce the TinyChange benchmark as a way to measure the sensitivity of audit methods in the context of small, realistic model changes.

Augustin Godinot, Erwan Le Merrer, Camilla Penzo et al.

The deployment of machine learning models in operational contexts represents a significant investment for any organisation. Consequently, the risk of these models being misappropriated by competitors needs to be addressed. In recent years, numerous proposals have been put forth to detect instances of model stealing. However, these proposals operate under implicit and disparate data and model access assumptions; as a consequence, it remains unclear how they can be effectively compared to one another. Our evaluation shows that a simple baseline that we introduce performs on par with existing state-of-the-art fingerprints, which, on the other hand, are much more complex. To uncover the reasons behind this intriguing result, this paper introduces a systematic approach to both the creation of model fingerprinting schemes and their evaluation benchmarks. By dividing model fingerprinting into three core components -- Query, Representation and Detection (QuRD) -- we are able to identify $\sim100$ previously unexplored QuRD combinations and gain insights into their performance. Finally, we introduce a set of metrics to compare and guide the creation of more representative model stealing detection benchmarks. Our approach reveals the need for more challenging benchmarks and a sound comparison with baselines. To foster the creation of new fingerprinting schemes and benchmarks, we open-source our fingerprinting toolbox.

Rachid Guerraoui, Rafael Pinot, Geovani Rizk et al.

Batch normalization has proven to be a very beneficial mechanism to accelerate the training and improve the accuracy of deep neural networks in centralized environments. Yet, the scheme faces significant challenges in federated learning, especially under high data heterogeneity. Essentially, the main challenges arise from external covariate shifts and inconsistent statistics across clients. We introduce in this paper Federated BatchNorm (FBN), a novel scheme that restores the benefits of batch normalization in federated learning. Essentially, FBN ensures that the batch normalization during training is consistent with what would be achieved in a centralized execution, hence preserving the distribution of the data, and providing running statistics that accurately approximate the global statistics. FBN thereby reduces the external covariate shift and matches the evaluation performance of the centralized setting. We also show that, with a slight increase in complexity, we can robustify FBN to mitigate erroneous statistics and potentially adversarial attacks.

Sayan Biswas, Davide Frey, Romaric Gaudel et al.

Decentralized learning (DL) is an emerging paradigm of collaborative machine learning that enables nodes in a network to train models collectively without sharing their raw data or relying on a central server. This paper introduces Zip-DL, a privacy-aware DL algorithm that leverages correlated noise to achieve robust privacy against local adversaries while ensuring efficient convergence at low communication costs. By progressively neutralizing the noise added during distributed averaging, Zip-DL combines strong privacy guarantees with high model accuracy. Its design requires only one communication round per gradient descent iteration, significantly reducing communication overhead compared to competitors. We establish theoretical bounds on both convergence speed and privacy guarantees. Moreover, extensive experiments demonstrating Zip-DL's practical applicability make it outperform state-of-the-art methods in the accuracy vs. vulnerability trade-off. Specifically, Zip-DL (i) reduces membership-inference attack success rates by up to 35% compared to baseline DL, (ii) decreases attack efficacy by up to 13% compared to competitors offering similar utility, and (iii) achieves up to 59% higher accuracy to completely nullify a basic attack scenario, compared to a state-of-the-art privacy-preserving approach under the same threat model. These results position Zip-DL as a practical and efficient solution for privacy-preserving decentralized learning in real-world applications.

Augustin Godinot, Gilles Tredan, Erwan Le Merrer et al.

Auditors need robust methods to assess the compliance of web platforms with the law. However, since they hardly ever have access to the algorithm, implementation, or training data used by a platform, the problem is harder than a simple metric estimation. Within the recent framework of manipulation-proof auditing, we study in this paper the feasibility of robust audits in realistic settings, in which models exhibit large capacities. We first prove a constraining result: if a web platform uses models that may fit any data, no audit strategy -- whether active or not -- can outperform random sampling when estimating properties such as demographic parity. To better understand the conditions under which state-of-the-art auditing techniques may remain competitive, we then relate the manipulability of audits to the capacity of the targeted models, using the Rademacher complexity. We empirically validate these results on popular models of increasing capacities, thus confirming experimentally that large-capacity models, which are commonly used in practice, are particularly hard to audit robustly. These results refine the limits of the auditing problem, and open up enticing questions on the connection between model capacity and the ability of platforms to manipulate audit attempts.

Aurélien Bellet, Edwige Cyffers, Davide Frey et al.

Decentralized Learning (DL) enables users to collaboratively train models without sharing raw data by iteratively averaging local updates with neighbors in a network graph. This setting is increasingly popular for its scalability and its ability to keep data local under user control. Strong privacy guarantees in DL are typically achieved through Differential Privacy (DP), with results showing that DL can even amplify privacy by disseminating noise across peer-to-peer communications. Yet in practice, the observed privacy-utility trade-off often appears worse than in centralized training, which may be due to limitations in current DP accounting methods for DL. In this paper, we show that recent advances in centralized DP accounting based on Matrix Factorization (MF) for analyzing temporal noise correlations can also be leveraged in DL. By generalizing existing MF results, we show how to cast both standard DL algorithms and common trust models into a unified formulation. This yields tighter privacy accounting for existing DP-DL algorithms and provides a principled way to develop new ones. To demonstrate the approach, we introduce MAFALDA-SGD, a gossip-based DL algorithm with user-level correlated noise that outperforms existing methods on synthetic and real-world graphs.

José Suárez-Varela, Miquel Ferriol-Galmés, Albert López et al.

During the last decade, Machine Learning (ML) has increasingly become a hot topic in the field of Computer Networks and is expected to be gradually adopted for a plethora of control, monitoring and management tasks in real-world deployments. This poses the need to count on new generations of students, researchers and practitioners with a solid background in ML applied to networks. During 2020, the International Telecommunication Union (ITU) has organized the "ITU AI/ML in 5G challenge'', an open global competition that has introduced to a broad audience some of the current main challenges in ML for networks. This large-scale initiative has gathered 23 different challenges proposed by network operators, equipment manufacturers and academia, and has attracted a total of 1300+ participants from 60+ countries. This paper narrates our experience organizing one of the proposed challenges: the "Graph Neural Networking Challenge 2020''. We describe the problem presented to participants, the tools and resources provided, some organization aspects and participation statistics, an outline of the top-3 awarded solutions, and a summary with some lessons learned during all this journey. As a result, this challenge leaves a curated set of educational resources openly available to anyone interested in the topic.

Alex Auvolat, Yérom-David Bromberg, Davide Frey et al.

Recent works have proposed new Byzantine consensus algorithms for blockchains based on epidemics, a design which enables highly scalable performance at a low cost. These methods however critically depend on a secure random peer sampling service: a service that provides a stream of random network nodes where no attacking entity can become over-represented. To ensure this security property, current epidemic platforms use a Proof-of-Stake system to select peer samples. However such a system limits the openness of the system as only nodes with significant stake can participate in the consensus, leading to an oligopoly situation. Moreover, this design introduces a complex interdependency between the consensus algorithm and the cryptocurrency built upon it. In this paper, we propose a radically different security design for the peer sampling service, based on the distribution of IP addresses to prevent Sybil attacks. We propose a new algorithm, $\scriptstyle{BASALT}$, that implements our design using a stubborn chaotic search to counter attackers' attempts at becoming over-represented. We show in theory and using Monte Carlo simulations that $\scriptstyle{BASALT}$ provides samples which are extremely close to the optimal distribution even in adversarial scenarios such as tentative Eclipse attacks. Live experiments on a production cryptocurrency platform confirm that the samples obtained using $\scriptstyle{BASALT}$ are equitably distributed amongst nodes, allowing for a system which is both open and where no single entity can gain excessive power.

George Giakkoupis, Anne-Marie Kermarrec, Olivier Ruas et al.

K-Nearest-Neighbors (KNN) graphs are central to many emblematic data mining and machine-learning applications. Some of the most efficient KNN graph algorithms are incremental and local: they start from a random graph, which they incrementally improve by traversing neighbors-of-neighbors links. Paradoxically, this random start is also one of the key weaknesses of these algorithms: nodes are initially connected to dissimilar neighbors, that lie far away according to the similarity metric. As a result, incremental algorithms must first laboriously explore spurious potential neighbors before they can identify similar nodes, and start converging. In this paper, we remove this drawback with Cluster-and-Conquer (C 2 for short). Cluster-and-Conquer boosts the starting configuration of greedy algorithms thanks to a novel lightweight clustering mechanism, dubbed FastRandomHash. FastRandomHash leverages random-ness and recursion to pre-cluster similar nodes at a very low cost. Our extensive evaluation on real datasets shows that Cluster-and-Conquer significantly outperforms existing approaches, including LSH, yielding speed-ups of up to x4.42 while incurring only a negligible loss in terms of KNN quality.

Daniel Bosk, Yérom-David Bromberg, Sonja Buchegger et al.

Mass surveillance of the population by state agencies and corporate parties is now a well-known fact. Journalists and whistle-blowers still lack means to circumvent global spying for the sake of their investigations. With Spores, we propose a way for journalists and their sources to plan a posteriori file exchanges when they physically meet. We leverage on the multiplication of personal devices per capita to provide a lightweight, robust and fully anonymous decentralised file transfer protocol between users. Spores hinges on our novel concept of e-squads: one's personal devices, rendered intelligent by gossip communication protocols, can provide private and dependable services to their user. People's e-squads are federated into a novel onion routing network, able to withstand the inherent unreliability of personal appliances while providing reliable routing. Spores' performances are competitive, and its privacy properties of the communication outperform state of the art onion routing strategies.

Loïck Bonniot, Christoph Neumann, François Taïani

Diagnosing problems in Internet-scale services remains particularly difficult and costly for both content providers and ISPs. Because the Internet is decentralized, the cause of such problems might lie anywhere between an end-user's device and the service datacenters. Further, the set of possible problems and causes is not known in advance, making it impossible in practice to train a classifier with all combinations of problems, causes and locations. In this paper, we explore how different machine learning techniques can be used for Internet-scale root cause analysis using measurements taken from end-user devices. We show how to build generic models that (i) are agnostic to the underlying network topology, (ii) do not require to define the full set of possible causes during training, and (iii) can be quickly adapted to diagnose new services. Our solution, DiagNet, adapts concepts from image processing research to handle network and system metrics. We evaluate DiagNet with a multi-cloud deployment of online services with injected faults and emulated clients with automated browsers. We demonstrate promising root cause analysis capabilities, with a recall of 73.9% including causes only being introduced at inference time.

Davide Frey, Marc X. Makkes, Pierre-Louis Roman et al.

Blockchains have a storage scalability issue. Their size is not bounded and they grow indefinitely as time passes. As of August 2017, the Bitcoin blockchain is about 120 GiB big while it was only 75 GiB in August 2016. To benefit from Bitcoin full security model, a bootstrapping node has to download and verify the entirety of the 120 GiB. This poses a challenge for low-resource devices such as smartphones. Thankfully, an alternative exists for such devices which consists of downloading and verifying just the header of each block. This partial block verification enables devices to reduce their bandwidth requirements from 120 GiB to 35 MiB. However, this drastic decrease comes with a safety cost implied by a partial block verification. In this work, we enable low-resource devices to fully verify subchains of blocks without having to pay the onerous price of a full chain download and verification; a few additional MiB of bandwidth suffice. To do so, we propose the design of diet nodes that can securely query full nodes for shards of the UTXO set, which is needed to perform full block verification and can otherwise only be built by sequentially parsing the chain.