Privacy-Preserving Deep Learning: Revisited and Enhanced

By Le Trieu Phong, Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai. 2017

206 citations recorded by the source corpus.

Subfields: Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Safety, Robustness & Interpretability

Citations in the corpus

The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks (2018, background): “Other related work [41] studies membership attacks on differentially private training, although in the setting of a distributed honest-but-curious server.”
Exploiting Unintended Feature Leakage in Collaborative Learning (2018, background): “[49] show that, in the collaborative deep learning protocol of [52], an honest-but-curious server can partially recover participants’ training inputs from their gradient updates under the (greatly simplified) assumption that the batch consists of a single input.”
Beyond Inferring Class Representatives: User-Level Privacy Leakage From Federated Learning (2018, methodology): “By contrast, the reconstruction attack aims to construct the training samples by accessing a learning model [3], [4], [16].”
A Simple and Efficient Federated Recommender System (2019, methodology): “Encryption techniques are applied in [3, 6] to prevent an honest-but-curious parameter server from reconstructing users’ gradients or parameters updates.”
A Survey of Privacy Attacks in Machine Learning (2020, background): “While this did not impact the model performance, it was later shown to be an inadequate measure [76].”
GAN-Leaks: A Taxonomy of Membership Inference Attacks against Generative Models (2019, background): “el was trained on personal data without an adequate legal basis [5]. Membership inference against discriminative deep learning 1 arXiv:1909.03935v1 [cs.LG] 9 Sep 2019 models has been largely explored [9,13,21,43,57,60,73,88], while inference against generative models is still an open question. This ”
Evaluating Gradient Inversion Attacks and Defenses in Federated Learning (2021, background): “[Phong et al., 2017] brings theoretical insights on this task by proving that such reconstruction is possible with a single-layer neural network.”
Privacy preservation in federated learning: An insightful survey from the GDPR perspective (2020, methodology): “que. A malicious coordination server can also recover partial personal data by inspecting the proportionality between locally trained model parameters sent to the server and the original data samples [12], [99]. Reconstruction attacks using MI and GANs are only feasible if and only if all class memb”
The Secret Sharer: Measuring Unintended Neural Network Memorization & Extracting Secrets (2018, background): “enses is often differential privacy [1,7,13,15,16]. Our analysis in Section9.3directly follows this line of work and we conﬁrm that it empirically prevents the exposure of secrets. Other related work [41] studies membership attacks on differentially private training, although in the setting of a dis”
Federated Learning on Non-IID Data: A Survey (2021, background): “from the gradient information of both shallow and deep neural networks [2, 107, 134, 165, 172].”
An Overview of Federated Deep Learning Privacy Attacks and Defensive Strategies (2020, methodology): “Such a method was proposed by [16] to secure DSGD but may easily be adapted to work on federated averaging.”
Eavesdrop the Composition Proportion of Training Labels in Federated Learning (2019, background): “relatively less-studied. In [26], the authors assume that the central server is malicious and train a GAN to reproduce data samples similar to the privacy of clients. Similar idea can also be seen in [27]. Please see SectionVIfor a more comprehensive discussion on related works. These previously-stu”
Inference Attacks Against Collaborative Learning (2018, background): “m generating an input that belongs to a particular class or even learning what such an input looks like.” It is not clear to us why this property is desirable or whether it is achievable. Aono et al. [2] show that, in the collaborative deep learning protocol of [44], an honest-but-curious server can”
Inverting Gradients - How easy is it to break privacy in federated learning? (2020, background): “Recovery of image data from gradient information was first discussed in [25, 24] for neural networks, who prove that recovery is possible for a single neuron or linear layer.”
A Framework for Evaluating Gradient Leakage Attacks in Federated Learning (2020, background): “on-phase privacy attacks. Gradient leakage attacks, formulated in CPL of dierent forms, or those in literature [10,45,43], are one type of privacy exploits in training phase. In addition, Aono et al [2,1] proposed a privacy attack, which partially recovers some private data of the clients based on t”
Deep Learning in Mobile and Wireless Networking: A Survey (2018, background): “Their framework is revisited and improved in [429], where another group of researchers employ”