acceptodds
Withdrawn from ICLR 2027

Fair Bandits Under Privacy and Adversarial Constraints

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Differential privacy (DP) is widely used in contextual bandits to protect sensitive user data. In this paper, we show how an adversary can conceal fairness-targeted corruptions within DP noise while overcoming its smoothing effect on reward estimates, which otherwise makes non-smooth reward functions less sensitive to small perturbations. The attack can maximize group regret disparity under a KL budget relative to the DP noise. We analyze four methods for spending the attack budget: Variant I (diffuse) spreads the perturbation uniformly across all dimensions, Variant II (concentrated) aligns the perturbation with the group-boundary normal, Variant III (burst) concentrates the budget into a short temporal window, and Variant IV (adaptive) conditions the perturbation direction on the observed noisy context. Experiments on six bandit algorithms across three real-world datasets show that the concentrated attack causes the largest fairness damage at a modest regret cost, and DP with per-dimension monitoring protects group fairness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential privacy, contextual bandits, algorithmic fairness, adversarial robustness, certified smoothing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_goGwOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper looks at context-manipulation attacks against local differentially private bandits. The attacker's goal is to slightly chang the private contexts to make learning performance (regret) worse for a target group, but keeps the changes within a KL divergence budget so it just looks like normal noise. It compares four attack methods on bandit datasets.

Critical strengths

n/a

Critical weaknesses
  1. The attacker's objective is given in Eq. (5). However, in the later theorems, I don't see guarantees for exactly the same objective. Am I missing anything?
  2. In table 1, how come the Rel. ↑ (II) for NeuralLinUCB is 20.520.5\\%? isn't it (22.9−15.6)/15.6=46.7(22.9-15.6)/15.6 = 46.7\\%?
Questions for authors

please see weakness

Reviewer_5ZfTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the problem of fairness-target contextual bandits with differential privacy on context under an adversarial environment. The authors consider group regret disparity under a KL budget relative to the DP noise. They propose Group-Targeted Context Manipulation Attack (GTCMA). Then analyze it as an optimization problem. They consider four cases of attack budget cost. The four attack variants differ in how the budget is allocated: Variant I distributes it evenly over all d coordinates, Variant II places it along the normal of the group boundary, Variant III spends it within a brief window of rounds rather than across the horizon, and Variant IV chooses the direction as a function of the privatized context it observes. Finally, the authors run some experiments to verify their methods.

Critical strengths

The problem studied in the paper is interesting.

Critical weaknesses
  1. The theoretical results are trivial and weird. For example, the result of Proposition 2 follows from Assumption 2 about the learner rather than from analysis.

  2. The experimental data have many error concerns. The data in Table 1 have some mistakes concern. For example, the "Rel. ↑ (II)" for NeuralLinUCB should be around 46.8% if the final column is the relative increase under Attack II over the no-attack baseline. But the table shows the value as "20.5%".

Questions for authors
  1. Could you provide upper bounds or lower bounds on the group regret you defined in (3) and show the dependence on TT, KK or privacy parameters?

  2. What is the real-world motivation for your problem of "adversarial fairness for DP-protected contextual bandits"?

Reviewer_TYgwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies an adversarial fairness attack on DP contextual bandits where group membership is inferred from features rather than observed. The learner cannot separate the attack from the privacy noise and the attacker’s goal is not aggregate regret but group-level regret disparity. The authors therefore proposer GTCMA that maximize group regret disparity. They formulate the attack as a KL-divergence budget-allocation problem to ensure the corruption remains statistically hidden within the DP noise. They distinguish two channels: a reward channel and a group channel. Experiments on different bandit algorithms across three datasets show that the concentrated attack is the most damaging. And raising disparity far more than regret.

Critical strengths
  1. They identify that in the fairness-under-unawareness setting, DP noise does two things at once: it hides the perturbation and it smooths the reward. The authors split this into two channels, show that the reward channel is provably bounded while the group channel is not, also confirm experimentally that privacy noise offers no protection on the group channel. It shows a previously not formulated perpective between deploying DP for privacy and preserving group fairness.
  2. The four attack variants map onto where to spend and when to spend, each with a matching theorem, and the experiments verify the predictions. Also, the authors mark some conditions results and some assumptions, showing agreement between theory and experiment.
Critical weaknesses
  1. Theorem 2 bounds the noise-averaged auxiliary function, but the learner never computes this average and sees noisy realization per round. Although authors acknowledge these in Remark 1 and introduce eTe_T, which is not bounded. The idea of this paper is " reward channel bounded, group channel vulnerable" but newly introduced eTe_T has not been verified.
  2. The effectiveness of the concentrated attack (Variant II) relies heavily on the adversary possessing an accurate estimate of the unit normal ww of the boundary separating the demographic groups. In a realistic online scenario, the learner's decision boundaries are updated based on feedback and are often non-linear, which leaves questions of practical.
Questions for authors
  1. Is it possible to bound the constant eTe_T in Assumption 2 for at least one concrete learner (e.g. DP-LinUCB)? Without this, the reward-channel robustness claim rests on an unverified assumption rather than a proven property.
  2. On German Credit, which has the largest d/s and so the largest predicted concentration advantage, the concentrated and adaptive variants are closer than the spatial theory predicts (Figure 5). How to explain this?
  3. Section 2.2 defines regret ΔR\Delta R, but the experiments report ΔDP\Delta DP.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.