Federated Boosted Decision Trees with Differential Privacy

Maddock, Samuel; Cormode, Graham; Wang, Tianhao; Maple, Carsten; Jha, Somesh

doi:10.1145/3548606.3560687

Cited by 14 publications

(14 citation statements)

References 40 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This enables us to design an efficient privacy auditing method for federated learning. The closest related work is CANIFE [Maddock et al, 2022], whose goal is also to perform privacy auditing of federated learning models. Compared to CANIFE, our technique requires minimal assumptions on adversarial knowledge (as the adversary crafts random canaries for poisoning model updates), and are effective for a range of adversarial models, including realistic adversaries with access only to the final model.…”

Section: Discussionmentioning

confidence: 99%

“…A canary update that was not observed in training will have vanishing cosine with the final model, while the cosine with one that was observed will be positive. Prior work has attempted to craft a canary example/update by finding an example/vector that is orthogonal to the clean data update [Maddock et al, 2022;Nasr et al, 2021]. Such approaches require knowledge of the model gradients, but we argue this is unnecessary for high dimensional models, because all vectors are nearly orthogonal to each other with high probability.…”

Section: Privacy Estimation For Fl With Random Canariesmentioning

confidence: 99%

“…For example, a common approach is to craft a "canary" training example on which the membership is being tested, which typically requires an adversary to have access to the underlying dataset and knowledge of the domain and model architecture. Finally, such techniques typically grant the adversary unrealistic power, for example (and in particular) the ability to inspect all model iterates during training [Maddock et al, 2022], something which may or may not be reasonable depending on the system release model.…”

Section: Introductionmentioning

confidence: 99%

“…Most privacy auditing techniques [Jagielski et al, 2020;Nasr et al, 2021;Lu et al, 2022;Zanella-Béguelin et al, 2022] have been designed for centralized settings, with the exception of CANIFE [Maddock et al, 2022], suitable for privacy auditing of federated learning deployments. CANIFE operates under a strong adversarial model, assuming knowledge of all intermediary model updates, as well as local model updates sent by a subset of clients in each round of training.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

One-shot Empirical Privacy Estimation for Federated Learning

Andrew¹,

Kairouz²,

Oh³

et al. 2023

Preprint

View full text Add to dashboard Cite

Privacy auditing techniques for differentially private (DP) algorithms are useful for estimating the privacy loss to compare against analytical bounds, or empirically measure privacy in settings where known analytical bounds on the DP loss are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g., knowledge of intermediate model iterates or the training data distribution), are tailored to specific tasks and model architectures, and require retraining the model many times (typically on the order of thousands). These shortcomings make deploying such techniques at scale difficult in practice, especially in federated settings where model training can take days or weeks. In this work, we present a novel "one-shot" approach that can systematically address these challenges, allowing efficient auditing or estimation of the privacy loss of a model during the same, single training run used to fit model parameters. Our privacy auditing method for federated learning does not require a priori knowledge about the model architecture or task. We show that our method provides provably correct estimates for privacy loss under the Gaussian mechanism, and we demonstrate its performance on a well-established FL benchmark dataset under several adversarial models.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Privacy Estimation For Fl With Random Canariesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

One-shot Empirical Privacy Estimation for Federated Learning

Andrew¹,

Kairouz²,

Oh³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…To analyze such data, deep learning algorithms may require complex neuron-based training with high overheads. Conversely, tree-based learning models show advantages in accurate performance and easy deployment [2,9]. To drive the ensemble tree-based model to play its advantages in FLbased attack detection service on IoT edge, there is still a lack of discussion and solutions to efficiently coordinate the tree-based FL design with the in-network ML-based attack detection in distributed IoT scenario.…”

Section: Introductionmentioning

confidence: 99%

Federated Learning-Based In-Network Traffic Analysis on IoT Edge

Zang

Zheng

Koziak

et al. 2023

2023 IFIP Networking Conference (IFIP Networking)

View full text Add to dashboard Cite

The rise of IoT-connected devices has led to an increase in collected data for service and traffic analysis, but also to emerging threats and attacks. In-network machine learningbased attack detection has proven effective in fast response, but scaling to distributed IoT edge devices risks increasing communication overheads and raising data privacy concerns. To address these concerns, we present FLIP4, a distributed innetwork attack detection framework based on federated tree models. FLIP4 maintains data privacy by enabling distributed machine learning training while keeping data local on IoT edge, and provides in-network inference within the programmable data plane on edge gateway for timely attack labeling and mitigation. Evaluation results show that FLIP4 can accurately detect attacks while maintaining source data privacy and enabling lightweight deployment on IoT edge.

show abstract

Random forest with differential privacy in federated learning framework for network attack detection and classification

Markovic,

Leon,

Buffoni

et al. 2024

Appl Intell

View full text Add to dashboard Cite

Communication networks are crucial components of the underlying digital infrastructure in any smart city setup. The increasing usage of computer networks brings additional cyber security concerns, and every organization has to implement preventive measures to protect valuable data and business processes. Due to the inherent distributed nature of the city infrastructures as well as the critical nature of its resources and data, any solution to the attack detection calls for distributed, efficient and privacy preserving solutions. In this paper, we extend the evaluation of our federated learning framework for network attacks detection and classification based on random forest. Previously the framework was evaluated only for attack detection using four well-known intrusion detection datasets (KDD, NSL-KDD, UNSW-NB15, and CIC-IDS-2017). In this paper, we extend the evaluation for attack classification. We also evaluate how adding differential privacy into random forest, as an additional protective mechanism, affects the framework performances. The results show that the framework outperforms the average performance of independent random forests on clients for both attack detection and classification. Adding differential privacy penalizes the performance of random forest, as expected, but the use of the proposed framework still brings benefits in comparison to the use of independent local models. The code used in this paper is publicly available, to enable transparency and facilitate reproducibility within the research community.

show abstract

Federated Boosted Decision Trees with Differential Privacy

Abstract: How to cite:Please refer to published version for the most recent bibliographic citation information. If a published version is known of, the repository item page linked to above, will contain details on accessing it.

Cited by 14 publications

References 40 publications

One-shot Empirical Privacy Estimation for Federated Learning

One-shot Empirical Privacy Estimation for Federated Learning

Federated Learning-Based In-Network Traffic Analysis on IoT Edge

Random forest with differential privacy in federated learning framework for network attack detection and classification

Contact Info

Product

Resources

About