SRATTA: Sample Re-ATTribution Attack of Secure Aggregation in Federated Learning.
Authors: Tanguy Marchand, Regis Loeb, Ulysse Marteau-Ferey, Jean Ogier Du Terrail, Arthur Pignet
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that SRATTA is both theoretically grounded and can be used in practice on realistic models and datasets. We also provide a Python implementation of SRATTA, and of the proposed defensive schemes. |
| Researcher Affiliation | Industry | Tanguy Marchand 1 Regis Loeb* 1 Ulysse Marteau-Ferey* 1 Jean Ogier du Terrail* 1 Arthur Pignet* 1 1Owkin Inc.. Correspondence to: Tanguy Marchand <tanguy.marchand@owkin.com>. |
| Pseudocode | Yes | Algorithm 1 Fed Avg, Algorithm 2 Local Update, Algorithm 3 SRATTA, Algorithm 4 Local Update Defended |
| Open Source Code | Yes | We also provide a Python implementation of SRATTA, and of the proposed defensive schemes. The code for SRATTAis available at https://github.com/owkin/SRATTA. |
| Open Datasets | Yes | Dataset used We perform SRATTA on four different datasets. Two of them are image datasets: CIFAR10 (Krizhevsky et al., 2009) and Fashion MNIST (Xiao et al., 2017). One is a binary dataset, the Primate Splice Junction Gene Sequences (hereafter DNA dataset) dataset available in the Open ML suite (Vanschoren et al., 2014). The final dataset is a multi-modal and multi-centric version of the TCGA-BRCA (Tomczak et al., 2015; Ogier du Terrail et al., 2022) dataset, containing binary, discrete and continuous entries. |
| Dataset Splits | No | The paper mentions 'training sequences' and 'test' data in various sections, but does not explicitly provide details about a 'validation' dataset split (e.g., percentages, sample counts, or specific methodology for validation). |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for its experiments; no specific GPU models, CPU types, or detailed computing environments are mentioned. |
| Software Dependencies | No | The paper mentions a 'Python implementation' but does not specify version numbers for Python itself or any other software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | G. Hyper-parameters used in numerical experiments Table 4 lists the hyper-parameters used in the different numerical experiments shown in this paper. Figure Dataset # centers #Dk # batch # hid. neur. nupdates tmax # trainings lr Figure 1 CIFAR10 5 100 8 1000 5 20 20 0.1 Figure 1 FMNIST 5 100 8 1000 5 20 20 0.5 Figure 1 DNA 5 100 8 1000 5 20 20 1.0 Figure 1 TCGA 5 100 8 1000 5 20 20 0.8 |