Our Data, Ourselves: Privacy Via Distributed Noise Generation

By Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, Moni Naor. 2006

2,443 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Electrical and Electronic Engineering, Privacy & Security

How researchers see this paper

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

In Artificial Intelligence

In Electrical and Electronic Engineering

In Privacy & Security

Citations in the corpus

Deep Learning with Differential Privacy (2016, background): “Differential privacy for repeated applications of additivenoise mechanisms follows from the basic composition theorem [19, 20], or from advanced composition theorems and their refinements [24, 34, 23, 11].”
Advances and Open Problems in Federated Learning (2019, background): “On the other hand, in the centralized setup where all the users’ data is already assumed to be held by a single entity and hence no communication is required, trade-offs between accuracy and privacy have been extensively studied in central DP starting with the foundational work of [147, 146].”
Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data (2016, background): “When teachers reach a strong quorum, this allows us to bound privacy costs more strictly.”
Practical and Private (Deep) Learning without Sampling or Shuffling (2021, methodology): “Assuming the data set D is sampled i.i.d. from a distribution τ , and the algorithm A outputs (cid:98) θ ∈ C , we want to minimize All the algorithms in this paper guarantee differential privacy [20, 21] and R´enyi differential privacy [53] (See Section 3 for details).”
The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation (2021, methodology): “The closest work to ours is cpSGD (Agarwal et al., 2018), which also serves as an inspiration for much of our work. cpSGD uses a distributed version of the binomial mechanism (Dwork et al., 2006a) to achieve distributed DP.”
Privacy Amplification by Decentralization (2020, methodology): “A popular approach is to resort to cryptographic primitives to securely aggregate user contributions [14, 31, 7, 10, 21].”
Differentially Private Learning Needs Better Features (or Much More Data) (2020, methodology): “…observation that model gradients lie in a low-dimensional subspace (Kairouz et al., 2020; Zhou et al., 2020b). satisfies (ε, δ)-differential privacy (Dwork et al., 2006a), if for any datasets D,D′ that differ in one record, and any set of models S: Pr[A(D) ∈ S] ≤ eε Pr[A(D′) ∈ S] + δ. DP bounds…”
On Distributed Differential Privacy and Counting Distinct Elements (2020, methodology): “In particular, [12] showed that ( ε, δ )-DP Setting δ = 2 − log k n for a suﬃciently large k and combining Lemma 20 and Theorem 2 gives us the desired lower bound against DP protocols.”
Training Production Language Models without Memorizing User Data (2020, background): “Differential privacy (DP) [DMNS06], [DKM + 06a] provides a gold standard for performing learning tasks over sensitive data.”
Privacy Amplification via Random Check-Ins (2020, background): “When such data is sensitive, steps must be taken to ensure privacy, and a formal guarantee of differential privacy (DP) [16, 15] is the gold standard.”
Private Counting from Anonymous Messages: Near-Optimal Accuracy with Vanishing Communication Overhead (2020, background): “1: procedure ANALYZERD 2: U ← number of messages received 3: return U − E[D] Privacy.”
Privacy in Deep Learning: A Survey (2020, background): “For ≥ 0, an algorithmA satisfies -DP [54, 20] if and only if for any pair of datasetsD andD′ that differ in only one element: P[A(D)= t]≤e P[A(D′)= t] ∀t (1)”
Pure Differentially Private Summation from Anonymous Messages (2020, methodology): “Since its introduction by Dwork et al. [DMNS06, DKM + 06], diﬀerential privacy (DP) has become widely popular as a rigorous mathematical deﬁnition of privacy.”
Theoretical Guarantees for Model Auditing with Finite Adversaries (2019, background): “This, in turn, led to the development of differential privacy (DP): a mathematically rigorous notion of privacy that prevents such memorization by placing an upper bound ε on the worst-case information loss [10–12].”
Rényi Differential Privacy of the Sampled Gaussian Mechanism (2019, methodology): “The notion of approximate differential privacy, which includes an additive δ term, appeared in the work by Dwork et al. [DKM + 06] in order to support analysis of the Gaussian noise mechanism.”
AdaCliP: Adaptive Clipping for Private SGD (2019, methodology): “We ﬁrst formally describe diﬀerential privacy [17] and the previous diﬀerentially private SGD methods.”
Differentially Private Learning with Adaptive Clipping (2019, neutral): “Differential privacy (DP) (Dwork et al., 2006a;b) is widely considered a gold standard for bounding and quantify-1 ing the privacy leakage of sensitive data when performing learning tasks.”