Truthful High Dimensional Sparse Linear Regression

We study the problem of fitting the high dimensional sparse linear regression model with sub-Gaussian covariates and responses, where the data are provided by strategic or self-interested agents (individuals) who prioritize their privacy of data disclosure. In contrast to the classical setting, our focus is on designing mechanisms that can effectively incentivize most agents to truthfully report their data while preserving the privacy of individual reports. Simultaneously, we seek an estimator which should be close to the underlying parameter. We attempt to solve the problem by deriving a novel private estimator that has a closed-form expression. Based on the estimator, we propose a mechanism which has the following properties via some appropriate design of the computation and payment scheme: (1) the mechanism is $(o(1), O(n^{-Ω({1})}))$-jointly differentially private, where $n$ is the number of agents; (2) it is an $o(\frac{1}{n})$-approximate Bayes Nash equilibrium for a $(1-o(1))$-fraction of agents to truthfully report their data; (3) the output could achieve an error of $o(1)$ to the underlying parameter; (4) it is individually rational for a $(1-o(1))$ fraction of agents in the mechanism; (5) the payment budget required from the analyst to run the mechanism is $o(1)$. To the best of our knowledge, this is the first study on designing truthful (and privacy-preserving) mechanisms for high dimensional sparse linear regression.

Paper

Similar papers

Peer review

Reviewer 81kj5/10 · confidence 2/52024-06-27

Summary

The authors present an $\varepsilon$-Bayesian incentive compatible and individually rational $k$-sparse linear regression algorithm with side payments for (almost all) privacy-oriented data providing agents. Remarkably, in the limit $d >> n >> \log d$, the accuracy and the needed budget both vanish.

Strengths

The paper tackles an interesting problem, at the heart of data collection for machine learning applications. The algorithms and the proof techniques are important advancements in the understanding of privacy-preserving machine learning with financial compensations to strategic data providers.

Weaknesses

The paper already has a lot of different notations, which are easily confusing for the reader. It is thus very important to significantly improve the writing, which is too often extremely hard on the reader: - Isn't the $\Pi_{\tau_\theta}$ the same operator as clipping (lines 2 and 3 of Algorithm 1)? It would make it much easier to read if all identical operators were written in the same way. - The partition $\hat D^0 \cup \hat D^1$ should be presented before its use in line 273. - Assumption 5 has a typo (I guess). It should be $D-i$ rather than $D_j$. There should also be a quantification over $i$ (unless each (c_i, D_i) is iid?). - The "symmetric threshold strategy" is used in the main text without definition (and without even a reference to the appendix!). The paper makes a common ground-truth assumption (same $\theta^*$). In practice, this may be unrealistic.

Questions

I don't see Assumption 3 as a generalization of Cummings et al. Given that $\delta \leq 1$, the assumption is essentially equivalent to (at least implied by) $f_i \leq 2c_i \varepsilon^3$, which is more demanding than $f_i \leq c_i \varepsilon^2$. It should be rather presented as a relaxation. Do the authors agree? I don't understand the claim (line 162) that "the one-round communication necessitates a closed-form estimator". Wouldn't any algorithm work? If so, I suggest that the authors merely present their new estimator as a proposal for covariance estimation, with a high-probability inversion guarantee (and even point to Theorem 3 as the end goal). I fail to understand how $\theta^*$ can be recovered without assumptions on the thresholds $r$, $\tau_x$, $\tau_y$ and $\lambda_n$. It seems that this should not be possible for arbitrary values of these parameters. Can the authors clarify this? Right now, though I think I mostly understand the algorithm, and its implications in terms of privacy, strategyproofness and individual rationality, its accuracy is a mystery to me, which prevents me from providing a larger score. I would be greatly thankful to the authors if they could help me gain insights into this aspect of their result.

Rating

5

Confidence

2

Soundness

2

Presentation

2

Contribution

4

Limitations

The authors should better stress that (i) Agents' "rationality" is merely on privacy leakage (they are indifferent to the trained model's behavior). (ii) Agents are assumed to have the same labeling function (parameter $\theta^*$). (iii) (Individually rational) agents whose data is too different are more likely removed from the system.

Authorsrebuttal2024-08-06

(Continued) Re: Q4

>**Q4: Right now, though I think I mostly understand the algorithm, and its implications in terms of privacy, strategyproofness and individual rationality, its accuracy is a mystery to me, which prevents me from providing a larger score. I would be greatly thankful to the authors if they could help me gain insights into this aspect of their results.** We thank the reviewer for appreciating our contributions. We present the accuracy result in Theorem 3. In the upper bound of Thm.3 there are two terms. The first one is $\tilde{O}(\frac{\alpha^2 n k}{\epsilon^2})$ (we assume the radius $r$ as a constant), where $\alpha$ is defined in Definition 6 is $1-\alpha$. Note that if all agents' cost is less than the threshold, then this term becomes zero. Moreover, as $1-\alpha$ means our participation goal, by choosing the parameter $\alpha$ sufficiently small (as suggested in Corollary 8), it will be dominated by the second term. The second term is $\tilde{O}(\frac{k}{n\epsilon^2})$, which is the cost due to privacy and statistical error. It is notable that even in the standard (non-private) high dimensional sparse linear model, the optimal estimation error is $\tilde{O}(\frac{k}{n})$, which only depends on $\log d$ and the sparsity. As we consider the high dimensional case where $n\ll d$, this bound will be $o(1)$. Thus, our second term is also very small and comparable to the non-private case.

Reviewer 81kj2024-08-09

I thank the authors for their response, which clarified several points. Overall, I believe that the paper would greatly gain from improved writing, which is why I stick with my rating. But I fully recognize the value of their contribution.

Authorsrebuttal2024-08-10

We appreciate the reviewer's recognition and feedback. For the camera-ready version, we will incorporate the requested clarifications and definitions on the additional page to ensure completeness. We will also correct any typos and unify the notation to enhance readability. If our response has addressed your concerns and positively influenced your view of our paper, we kindly ask you to consider revising the score.

Reviewer BxFH3/10 · confidence 4/52024-07-07

Summary

The paper solves the problem of high dimensional sparse regression with subgaussian covariates. Along with doing that they also ensure differential privacy of the data providers. Finally, they also provide a payment scheme that is individually rational and which incentivizes truthfulness.

Strengths

The paper provides an algorithm to solve the problem of high dimensional sparse regression with subgaussian covariates. They also make sure that the algorithm also ensures differential privacy of the data providers. Finally, they also provide a payment scheme that is individually rational and which incentivizes truthfulness.

Weaknesses

1) The paper seems to be a slight extension of differentially private logistic regression while also incorporating payments into it which does not seem like a major extension given the earlier works of Fallah et al. or Anjarlekar et al. have incorporated incentive-compatible payment mechanisms in traditional DP settings with Fallah et al. solving the problem for mean estimation while Anjarlekar et al. extending that work for the case of logistic regression. 2) The assumptions made in the paper do not seem practical (for example assumptions 3 and 4)

Questions

1) It is unclear to me why the authors have used Joint Differential Privacy instead of the original DP definition. Can the authors provide more justification for this? 2) Assumption 3 seems a bit vague. Can the authors clarify the specific structure of the upper bound used in the assumption? It seems more clear to have f(.) which is a monotonically increasing convex function of $\epsilon$. 3) Can the authors clarify more about the practical validity of Assumption 4? 4) A literature survey related to incorporating payments in differentially private models seems incomplete. Some relevant works are as follows [a] Justin Kang, Ramtin Pedarsani, & Kannan Ramchandran. (2024). The Fair Value of Data Under Heterogeneous Privacy Constraints in Federated Learning. [b] Ameya Anjarlekar, Rasoul Etesami, & R. Srikant. (2023). Striking a Balance: An Optimal Mechanism Design for Heterogenous Differentially Private Data Acquisition for Logistic Regression. 5) Can be approach be extended to scenarios incorporating heterogeneous differential privacy? Similarly can the proposed approach work for larger models such as deep neural networks? 6) It would have been better to add some experimental results to highlight how the payments and model error varies with a change in the differential privacy guarantees and other parameters.

Rating

3

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

NA

Authorsrebuttal2024-08-06

(Cont'd) Response to Reviewer BxFH' s questions

>**Q1: It is unclear to me why the authors have used Joint Differential Privacy instead of the original DP definition. Can the authors provide more justification for this?** Full differential privacy requires that all outputs by the mechanism, including the payment it allocates to a user, are insensitive to every user’s input. In our strategic user setting, the payment to every user is supposed to be kept secret. Therefore, it is more natural to assume that the payment $\pi_i$ to each user is only observable by the user $i$ while the estimate $\hat{\theta}$ is publicly observable. This consideration exactly falls into the motivation of Joint Differential Privacy. >**Q2: Assumption 3 seems a bit vague. Can the authors clarify the specific structure of the upper bound used in the assumption? It seems more clear to have $f(.)$ which is a monotonically increasing convex function of $\epsilon$.** The monotonicity in $\epsilon$ is intuitive: smaller values imply stronger privacy properties. Specifically, $\epsilon = 0 $ indicates the output is independent of user $i$’s data. Moreover, we use cubic term in $\epsilon$ to provide stronger constraints. **However, it should be noted that the bounding function $F = c_i (\delta+ 1) \epsilon^3$ is not associated with convexity. In our proof, we also did not use the convexity property.** >**Q3: Can the authors clarify more about the practical validity of Assumption 4?** In Assumption 4, we introduce the privacy cost coefficient $c_i$ as a random variable, sampled from some distribution. As mentioned earlier, the payment $\pi_i$ to each user is supposed to be private and the amount of payment is strongly linked to the privacy cost $c_i$. Therefore, it is very natural to assume that $c_i$ depends on each user's own data $D_i = (x_i,y_i)$ and conditioned on $D_i$, $c_i$ does not reveal any additional information on any other user's cost coefficient or data. Otherwise, a strategic user can use his privacy cost coefficient to infer other users' payments. >**Q4: A literature survey related to incorporating payments in differentially private models seems incomplete. Some relevant works are as follows [a] Justin Kang, Ramtin Pedarsani, Kannan Ramchandran. (2024). The Fair Value of Data Under Heterogeneous Privacy Constraints in Federated Learning. [b] Ameya Anjarlekar, Rasoul Etesami, R. Srikant. (2023). Striking a Balance: An Optimal Mechanism Design for Heterogenous Differentially Private Data Acquisition for Logistic Regression.** We thank the reviewer for providing the relevant literature to us. We will add them to the related work section. >**Q5: Can be approach be extended to scenarios incorporating heterogeneous differential privacy? Similarly can the proposed approach work for larger models such as deep neural networks?** Sadly, our current framework is not suitable for extending to other scenarios like heterogeneous differential privacy. This is because our privacy guarantee relies on the Billboard Lemma to broadcast the original DP to JDP, and it is unknown whether there is a heterogeneous version of Billboard Lemma. Meanwhile, designing high-dimensional truthful mechanisms with privacy constraints is an intrinsically hard problem. Our mechanism cannot be directly applied to larger models such as DNNs like many other truthful mechanisms. Given this, we believe that our work has been a leap and paved the way for future research. **Please see the above for response to question 6**

Authorsrebuttal2024-08-10

We thank the reviewer for their time and for mentioning the related work which we have not included. We will add these to our paper. If our response has addressed your concerns and positively influenced your view of our paper, we kindly ask you to consider revising the score.

Reviewer BxFH2024-08-12

I am still not convinced about the choice of the bounding function. Do the results still hold if we consider a general monotonic bounding function?

Authorsrebuttal2024-08-13

**Our upper bounding function has to include $c_i, \epsilon$ and $\delta$, where $c_i$ is the privacy cost coefficient.** | **Paper** | **Upper Bounding Function** | **Data Assumption** | **Privacy Guarantee** | |-----------|--------------------------------------|---------------------|-------------------------------| | Ours | $c_i(1+\delta) \epsilon^3 $ | Sub-Gaussian | $(\epsilon, \delta)$-JDP | | [1] |$ c_i(1+\gamma) \epsilon^4 $ | Sub-Gaussian | $(\epsilon, \gamma)$-RJDP | | [1] |$ c_i(1+\gamma) \epsilon^9 $ | Heavy-tailed | $(\epsilon, \gamma)$-RJDP | | [2] | $ c_i \epsilon^2 $ | Bounded |$ \epsilon$-JDP | It is clear that due to different data settings, our assumption is weaker than that of [1] but is stronger than the assumption in [2] which was first introduced by [3] and later adopted by [4]. To see why the upper bounding function in [2] is quadratic in $\epsilon$, we additionally define $g_i$ as the privacy cost for reporting $(x_i , y_i)$ to the mechanism $M$ when all other agents report $(X_{-i}, y_{−i})$ and the output is $\hat{\theta}$. In [3], the authors assume that $g_i\left(M, \hat{\theta},\left(x_i, y_i\right),\left(X_{-i}, y_{-i}\right)\right) \leq c_i \ln \left(\max_{y_i^{\prime}, y_i^{\prime \prime}} \frac{\operatorname{Pr}\left[M\left(X, y_i^{\prime}, y_{-i}\right)=\hat{\theta}\right]}{\operatorname{Pr}\left[M\left(X, y_i^{\prime \prime}, y_{-i}\right)=\hat{\theta}\right]}\right)$. This assumption is natural because if $g_i$ is small if the agent $i$'s report exerts little influence on $\hat{\theta}$ and therefore should be upper bounded by a function that depends on the effect that the agent $i$’s report has on the mechanism’s output. And by Composition Lemma (Lemma 3.2 in [3]), $\mathbb{E}\left[g_i\left(M, M(X, y),\left(x_i, y_i\right),\left(X_{-i}, y_{-i}\right)\right)\right]-\mathbb{E}\left[g_i\left(M, M\left(X, y_i^{\prime}, y_{-i}\right),\left(x_i, y_i\right),\left(X_{-i}, y_{-i}\right)\right)\right] \leq 2 c_i \epsilon\left(e^\epsilon-1\right) \leq 4 c_i \epsilon^2$ Therefore by setting the privacy cost function $f_i = \frac{1}{4} \mathbb{E}\left[g_i\left(M, M(X, y),\left(x_i, y_i\right),\left(X_{-i}, y_{-i}\right)\right)\right]$, it can be interpreted as the agent $i$’s expected cost for participating in the mechanism up to a scaling constant. Given that our assumption regarding $f_i$ stems from natural ground and has been used widely in similar settings, we argue that it is a reasonable assumption. [1] Qiu, Yuan, Jinyan Liu, and Di Wang. "Truthful Generalized Linear Models." arXiv preprint arXiv:2209.07815 (2022). [2] Cummings, Rachel, Stratis Ioannidis, and Katrina Ligett. "Truthful linear regression." Conference on Learning Theory. PMLR, 2015. [3] Chen, Yiling, et al. "Truthful mechanisms for agents that value privacy." ACM Transactions on Economics and Computation (TEAC) 4.3 (2016): 1-30. [4] Ghosh, Arpita, et al. "Buying private data without verification." Proceedings of the fifteenth ACM conference on Economics and computation. 2014.

Authorsrebuttal2024-08-14

Dear Reviewer BxFH, Please see above our detailed response to your question regarding our assumption 3. We believe that we have used a very general and practical bounding function. If you think our response clarifies your concerns, we kindly ask you to consider raising the score. Best Regards, Authors

Reviewer YR327/10 · confidence 3/52024-07-13

Summary

This paper focuses on mechanism design that incentivizes truthful data reporting while preserving privacy in the context of high dimensional sparse linear regression. The proposed mechanism is $(o(1),O(n^{-Omega(1)}))$-jointly differentially private, provides an estimator that is $o(1)$ accurate, is an approximate Bayes NE where most of the agents report truthfully, asymptotically individually rational, and requires a small payment budget.

Strengths

1. This paper has a very clean presentation despite the complicated problem setting and technical components. 2. The theoretical guarantees are comprehensive, including estimation error, privacy guarantee, truthfulness, individual rationality, and budget. 3. The private estimator is quite interesting.

Weaknesses

Currently, each agent only collects one data point. Is it possible to study the free-riding issue as well under the current framework?

Questions

See above

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

See above

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC