acceptodds
Under review as a conference paper at ICLR 2027

ADAN: Adaptive Direction-Aware Noise for Defending Iterative Gradient Inversion Attacks

Abstract

Iterative training data reconstruction is a class of gradient inversion attacks (GIAs) in which an adversary repeatedly updates a dummy input to align its gradient with the gradients of a target model, progressively recovering the ground-truth training data. Existing GIAs operate under a curious-but-honest setting. This setting provides the attackers with a significant advantage because they can utilize the model parameters to calculate the exact gradients during iterative optimization. By contrast, in this work, we consider a more challenging black-box setting in which the attacker has no access to model parameters. Based on this setting, we propose **Adaptive Direction-Aware Noise (ADAN)**, a mechanism that injects input-dependent, non-isotropic Gaussian noise into each queried gradient. ADAN employs a neural network to learn per-coordinate noise magnitudes adapted to the input data distribution, effectively steering the optimization geometry of iterative GIAs. Empirical results show that ADAN substantially outperforms isotropic gradient perturbation methods across multiple datasets and attack variants.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Learning-based defense, iterative gradient inversion attacks, input-adaptive noise, non-isotropic noise

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bAVAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose ADAN, which uses machine learning to learn the defense noise. The method adds Gaussian noise to the gradients, and the noise scale is learned. The authors train the noise scale by optimizing the extent an gradient-based attacker would deviate from the ground truth direction, while bounding the total noise within a given level to preserve training quality. Experiments shown better performance against isotropic noise, pruning, and clipping.

Critical strengths

The idea that the defense is aware that attacks use optimization methods is interesting, many attacks are based on this framework. This trained noise clearly beats isotropic noise in their experiments, claiming effectiveness of the idea.

Critical weaknesses
  1. The paper claims their defense to be coordinate-wise, but based on my understanding of the paper, the training objective (7) only depends on the noise scale hϕh_\phi on that sample. Moreover, in Appendix A.2 the noise scale is normalized, if that is added to the derivation of training objective then the training objective hϕh_\phi is constant, and the training objective is 0. In this case training does not yield useful information.
  2. The experiment scope is very limited, and the defense is only designed against attacks based on direct optimization of gradient difference. Methods not using optimization at all (e.g. R-GAP) or regularized objectives is also not considered. The defenses benchmarked against is also very weak, only plain isotropic noise, pruning, and clipping, while many adaptive and more advanced defenses exist.
Questions for authors
  1. Does the method outperform advanced defenses that are adaptive (e.g. CENSOR) and defenses that does not belong to the optimization framework?
  2. How do we obtain coordinate-wise defense noise if the target is a function of the total noise scale? My intuition of the target is that with high noise on data points and low noise on other points, the attacker minimizing loss will be directed to low-loss region, which is low noise and away from data points. This did not really utilize fact that the defense could drive the attacker the wrong way, and the current framework seems to be a different approach.
  3. Appendix A.2 states that the noise is normalized, but this is not included in section 4. How is the normalization applied during training and inference? If the total noise is fixed, equation 7 appears to be constant. Does this affect the training objective?
3 comments

Answer to Weakness 1.

Our method uses ADMM for constrained optimization. During optimization, the noise scale is not enforced as a hard constraint, so there is no possibility of the objective becoming invalid. Optimization proceeds in two directions: tangential and radial with respect to the noise. The noise-scale constraint restricts the radial direction, so the noise focuses more on optimizing its coordinate-wise proportions—i.e., the tangential direction. Under soft constraint, the final noise scale will not exactly equal the constraint value. Normalization is intended to force the noise scale to equal the constraint after the ADAN training by multiplying a scalar. Its purpose is to make the utility of the isotropic baseline and ADAN strictly comparable.

For more details, please refer to Answer to Questions.

Answer to Weakness 2.

Table 1. Unified quantitative comparison of defense performance across attacks. MSE ↑ indicates better privacy; Acc ↑ indicates better utility.

Attack (dataset)DefenseParamUtility (Acc %) ↑MSE ↑
DLG (CIFAR-10)None—61.70.043
DLG (CIFAR-10)Isotropicσ=0.0160.70.068
DLG (CIFAR-10)Isotropicσ=0.0558.20.115
DLG (CIFAR-10)Isotropicσ=0.1053.70.138
DLG (CIFAR-10)Isotropicσ=0.2049.10.159
DLG (CIFAR-10)ADANσ=0.0260.30.1588
DLG (CIFAR-10)ADANσ=0.0459.60.1869
DLG (CIFAR-10)ADANσ=0.0659.20.1844
DLG (CIFAR-10)ADANσ=0.0859.30.1789
DLG (CIFAR-10)Sparsifyp=0.0549.90.135
DLG (CIFAR-10)Sparsifyp=0.1048.50.114
DLG (CIFAR-10)Sparsifyp=0.2549.30.082
DLG (CIFAR-10)Sparsifyp=0.5048.90.058
DLG (CIFAR-10)Quantize2-bit48.00.177
DLG (CIFAR-10)Quantize4-bit48.10.170
DLG (CIFAR-10)Quantize8-bit48.70.170
DLG (CIFAR-10)ClipC=0.0549.90.042
DLG (CIFAR-10)ClipC=0.5049.30.042
DLG (CIFAR-10)ClipC=1.0049.00.042
GGL (ImageNet)None—85.30.082
GGL (ImageNet)Isotropicσ=0.0184.40.076
GGL (ImageNet)Isotropicσ=0.0581.00.082
GGL (ImageNet)Isotropicσ=0.1041.00.087
GGL (ImageNet)Isotropicσ=0.2021.30.100
GGL (ImageNet)ADANσ=0.0580.80.0724
GGL (ImageNet)ADANσ=0.1059.00.0813
GGL (ImageNet)ADANσ=0.1547.00.0849
GGL (ImageNet)ADANσ=0.2035.00.0870
GGL (ImageNet)Sparsifyp=0.1085.00.106
GGL (ImageNet)Sparsifyp=0.2584.30.095
GGL (ImageNet)Sparsifyp=0.5082.00.087
GGL (ImageNet)Quantize2-bit81.00.105
GGL (ImageNet)Quantize4-bit84.50.092
GGL (ImageNet)Quantize8-bit85.20.085
GGL (ImageNet)ClipC=0.1084.80.081
GGL (ImageNet)ClipC=0.5085.30.081
GGL (ImageNet)ClipC=1.0085.20.081
GGL (ImageNet)Soteriafixed80.80.0461
GIFD (CIFAR-10)Noneσ=0.0161.70.020
GIFD (CIFAR-10)Isotropicσ=0.0160.70.020
GIFD (CIFAR-10)ADANσ=0.0160.60.022
GIFD (CIFAR-10)Noneσ=0.0561.70.019
GIFD (CIFAR-10)Isotropicσ=0.0558.20.023
GIFD (CIFAR-10)ADANσ=0.0559.20.025
GIFD (CIFAR-10)Noneσ=0.1061.70.019
GIFD (CIFAR-10)Isotropicσ=0.1053.70.024
GIFD (CIFAR-10)ADANσ=0.1059.30.031
GIFD (CIFAR-10)Noneσ=0.2061.70.017
GIFD (CIFAR-10)Isotropicσ=0.2049.10.033
GIFD (CIFAR-10)ADANσ=0.2058.20.037
GIFD (ImageNet)Noneσ=0.0185.30.083
GIFD (ImageNet)Isotropicσ=0.0184.4†0.075
GIFD (ImageNet)ADANσ=0.0185.30.083
GIFD (ImageNet)Noneσ=0.0585.30.071
GIFD (ImageNet)Isotropicσ=0.0581.00.074
GIFD (ImageNet)ADANσ=0.0580.80.085
GIFD (ImageNet)Noneσ=0.1085.30.081
GIFD (ImageNet)Isotropicσ=0.1041.00.072
GIFD (ImageNet)ADANσ=0.1059.00.088
GIFD (ImageNet)Noneσ=0.2085.30.079
GIFD (ImageNet)Isotropicσ=0.2021.30.077
GIFD (ImageNet)ADANσ=0.2035.00.089

Analysis: On DLG/CIFAR-10, ADAN clearly lies in the upper-right region of the privacy–utility curve and outperforms the listed baselines. On GGL/ImageNet, ADAN outperforms Soteria and Isotropic, but it is not globally upper-right: Clip, Sparsify, and Quantize can match or exceed ADAN in the high-utility regime. Under GIFD, ADAN still outperforms None/Isotropic in most settings.

Additionally, as stated in the title, our proposed methods are designed for optimization-based GIA (SOTA). We will add GIFD, but R-GAP (closed-form GIA) is not included.

Answer to Question 1.

We agree that CENSOR achieves substantially higher utility than the other defense methods, but we do not consider a comparison with them to be fair. CENSOR avoids the utility–privacy trade-off by searching for a gradient that is orthogonal to the original gradient and maximizes utility. This is purchased by a very high computational cost: at each step the feasible set is the orthogonal complement of the original gradient, which is a (d−1)-dimensional set. Even if one restricts the search to a set of orthogonal basis directions, there are d−1 such directions, and evaluating them is already expensive. An exhaustive or naive search over the orthogonal complement is continuous and scales poorly with model dimension. A fair comparison should therefore account for this search cost, or compare methods under the same compute and privacy budget rather than utility alone.

Answer to Question 2.

The objective is not the total noise scale. The term related to the noise scale is the derivative of the noise’s L2 norm with respect to the input. Every coordinate of the noise is connected to the input in the computation graph, and the noise generator is built inside the leaked model. Therefore, when the attacker updates the dummy input by differentiating the gradient with respect to the input, backpropagation flows through the noise generator parameters. This derivative is affected by the entire ANAN model, which pushes the optimization of the dummy input in the opposite direction. This is the intuition behind why the defense works. See Fig. 4 for the detailed noise-generator design.

Normalizing the noise is only intended to compare privacy with the isotropic case under strictly equal utility. Normalization is multiplication by a required constant and is not included in the computation graph from the input to the noise.

Answer to Question 3.

Normalization is intended to force the noise scale to equal the constraint. Its purpose is to make the utility of the isotropic baseline and ADAN strictly comparable for comparison with the baseline, which is not a necessary step. This step must not affect the computation graph from the input to the noise; indeed, the term “normalization” is ambiguous here. What it means is that, after the noise is generated by ADAN, we multiply it by a constant chosen according to the desired strict constraint, and this constant does not affect the computation graph of ADAN. The pseudocode currently contains only the ADAN training, I will add pseudocode for the noise implementation later in the revised version.

Thank you to the authors for the response! Could the authors locate a figure version of the provided table in the revised manuscript? The tables are a bit hard to read and based on the description this does not seem to dominate other simple methods?

Q2. I understand that the objective is not the total noise scale, but my main concern is that the objective is a function of the noise scale. If the objective is a function of the noise scale, even if every coordinate do connect but their effect to the objective does not differ, they only learn through the aggregate noise scale. If so then the network only learned the scale of the noise, and not the direction of noise. After normalization the scale is normalized, and the direction remains without any training signal. This is also my main reason behind the rejection rating. Is any part of my understanding incorrect? If only the scale is learned then the direction used reduces to almost a random direction.

Q1. Even not for CENSOR specifically, there are definitely many baselines that both change the noise scale which you should compare with. (Take https://www.usenix.org/system/files/usenixsecurity25-xu-xiangrui.pdf for example). I haven't been following up on this topic for a while so I don't know the sota, but there definitely is thus a line of work using the SVD of the Jacobian to give a different noise direction. You can even test a random noise direction, the same noise changing class of defense is severely lacking.

Reviewer_f4TCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies defenses against iterative gradient inversion attacks (GIAs). Unlike the conventional honest-but-curious setting, where the attacker is assumed to have access to the model parameters, the authors consider a black-box setting in which these parameters are unavailable to the attacker. The authors propose an adaptive direction-aware noise mechanism, ADAN, which uses a neural network to predict coordinate-wise noise standard deviations from the input, producing input-dependent, anisotropic Gaussian perturbations to the gradients. The experimental results show that, under the black-box threat model considered in the paper, ADAN reduces reconstruction quality more effectively than isotropic noise.

Critical strengths
  • Mitigating the privacy leakage caused by gradient inversion attacks is an important problem in federated learning.
  • The proposed use of anisotropic gradient noise tailored to defending against GIAs is novel and empirically effective.
Critical weaknesses
  • The paper departs from the conventional threat model for gradient inversion attacks without sufficiently motivating this change. The paper does not sufficiently justify its black-box assumptions or provide a concrete federated learning scenario in which they hold. Moreover, withholding model parameters already restricts the attacker’s knowledge relative to the conventional honest-but-curious setting with access to these parameters. Thus, even if ADAN effectively mitigates GIAs in the proposed black-box setting, it does not establish its effectiveness against the privacy threats posed by attackers with access to model parameters.
  • The attack evaluation is limited. relying primarily on DLG and GGL. Although these are representative GIAs, the evaluation omits more recent methods, such as GIFD [1] and DGGI [2].
  • The coverage of defense baselines is also insufficient. The paper compares ADAN with isotropic noise, gradient clipping, and gradient pruning, but does not include defenses specifically designed to mitigate gradient inversion attacks, such as Soteria [3].

[1] GIFD: A Generative Gradient Inversion Method with Feature Domain Optimization [2] DGGI: Deep Generative Gradient Inversion with diffusion model [3] Soteria: Provable Defense Against Privacy Leakage in Federated Learning From Representation Perspective

Questions for authors
  • The paper focuses on a single local update step, whereas federated learning can involve multiple local updates. Does ADAN remain applicable in this setting?
  • Can the authors identify an existing federated learning framework that matches the threat model considered in this paper?
2 comments

Answer to Weakness 1.

The core setting of our threat model is a gradient-oracle interface: the model parameters are hidden or encrypted on a trusted server, while an attacker can repeatedly observe leaked, possibly perturbed, gradients.

Such a setting arises naturally in several practical scenarios. One representative case is distributed learning, such as federated learning (FL), where an attacker compromises the communication channel (I/O) of a client and thereby gains the ability to inject crafted inputs while observing the corresponding leaked gradients. Another example arises in machine-learning-as-a-service (MLaaS), where model providers keep model parameters hidden or encrypted to protect intellectual property, while still exposing gradient information to support downstream tasks such as data valuation, data selection, and data cleaning.

Why can't the attacker obtain model parameters? We consider the scenario where the local execution environment exposes only a narrow, well-defined interface: the client receives an encrypted model bundle from the server, which is decrypted within a trusted execution environment (TEE) or secure enclave (e.g., Intel SGX), and the decrypted weights are never materialized in accessible GPU memory. The attacker's hook intercepts gradient tensors flowing through the runtime after the forward-backward pass, but the model weights themselves remain protected within the enclave boundary. This architecture, encrypting model distribution while exposing gradient computation pipelines, is standard in production FL deployments (e.g., Google's FL infrastructure).

Answer to Weakness 2.

We now evaluate ADAN against multiple attack methods spanning all major categories:

Table 1. Generator-based attacks — GIFD (CIFAR10, pretrained DCGAN):

σNo DefenseIsotropicADAN
0.010.0200.0200.022
0.050.0190.0230.025
0.100.0190.0240.031
0.200.0170.0330.037

Table 2. Generator-based attacks — GIFD (ImageNet, BigGAN):

σNo DefenseIsotropicADAN
0.010.0830.0750.083
0.050.0710.0740.085
0.100.0810.0720.088
0.200.0790.0770.089

Analysis: ADAN provides much better protection against new generator-based attack baselines.

Answer to Weakness 3.

On GGL/ImageNet, Soteria uses a fixed configuration and reports an MSE of 0.0461. In contrast, our ADAN achieves MSE values of 0.0709, 0.0837, 0.0826, and 0.0891 at σ = 0.05, 0.10, 0.15, and 0.20, respectively—all higher than Soteria, indicating stronger reconstruction disruption and better defense. More importantly, this gain in defense strength does not come at the cost of model utility: even at our highest noise level (σ = 0.20), ADAN still maintains a utility of 85.3, compared with only 80.8 for Soteria. Therefore, on GGL/ImageNet, ADAN not only consistently outperforms Soteria’s fixed defense result in MSE but also preserves higher utility at high σ, achieving a better defense–utility trade-off. ADAN dominates the upper-right region of the privacy–utility curve: compared with baselines, ADAN can achieve better utility under approximately the same attack MSE.

Answer to Question 1.

Table 3. ADAN Defense under multiple local updates on CIFAR-10. Reconstruction MSE (↑ better privacy):

Local UpdateN=5 (MSE ↑)N=10 (MSE ↑)N=15 (MSE ↑)
DLG attack0.17570.18430.1912
GGL attack0.09310.10900.1005

Analysis: We clarify that multi-step local updates are inherently favorable to the defender and make the attack harder. Moreover, increasing this interval does not affect ADAN's training, because ADAN itself is a defense applied on top of local training. Under our threat model the local client is a black box to the attacker. For this reason, a larger interval just makes the defense stronger.

Answer to Question 2.

Please refer to Answer to Weakness 1.

I thank the authors for addressing some of my concerns, but I still have one concern. Although the gradient-oracle setting considered in this paper is meaningful and worth studying, existing GIAs typically assume a curious-but-honest setting where the attacker has access to model parameters. In contrast, the proposed defense operates under a setting that is more favorable to the defender. As a result, the privacy risks in the curious-but-honest setting commonly considered in GIA research may remain unaddressed by this defense. This leaves me concerned about the paper's positioning. I therefore keep my score.

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

This paper proposes ADAN (Adaptive Direction-Aware Noise) for defending iterative gradient inversion attacks. Gradient inversion attacks try to reconstruct private training data and their works proposes a defense but only in the black-box setting. Their proposed method is a learning based and non-isotropic perturbation method that injects input dependent noise into queried gradient. By changing the optimization, the proposed method defends against GIA attacks. They perform experiments using the CIFAR-10 and ImageNet datasets across federated learning settings. They show that their method performs well by stopping data reconstruction compared to different baselines.

Critical strengths

Here are some strengths of the paper:

  1. The paper identifies and tries to solve an important problem in the distributed learning setting where an attacker can gain access to sensitive private data

  2. The paper outlines the problem and compares against different methods and shows that their proposed defense works well

  3. They perform experiments on both small scale (CIFAR-10) and large scale (ImageNet) image datasets

Critical weaknesses

Here are some weaknesses of the paper in my opinion:

  1. My main problem with this paper is that it only works (shown to work) against black-box setting. The authors clearly outline that their method is proposed for that and that white-box setting is unlikely. But this assumption makes the proposed defense seem very weak. What if the attacker gains access to the parameters of the model? For a defense method to work well, it should be evaluated against every scenario including white-box settings. In fact, a good defense on white-box setting will perform well also on black-box if evaluated correctly as white-box attack settings are the strongest attacks. Can the authors evaluate their method on white-box setting? Can it work on that setting?

  2. I believe your defense method is stochastic in nature. Do you evaluate by using EoT (Expectation over Transformation) by averaging multiple runs?

Questions for authors

Please refer to the weaknesses section

2 comments

Answer to Weakness 1.

The reason our paper considers the black-box setting is that ADAN’s defense mechanism is applied to the parameters. A white-box setting cannot be considered, since that would amount to assuming the attacker has already bypassed the defense. This does not mean, however, that black-box is an easier setting than white-box. White-box assumes that the attacker has compromised the server, whereas black-box assumes that the attacker has compromised the client (shown in Figure 1). In attack scenarios, the two are on equal footing, they differ only in their manifestation.

Why can't the attacker obtain model parameters? We consider the scenario where the local execution environment exposes only a narrow, well-defined interface: the client receives an encrypted model bundle from the server, which is decrypted within a trusted execution environment (TEE) or secure enclave (e.g., Intel SGX), and the decrypted weights are never materialized in accessible GPU memory. The attacker's hook intercepts gradient tensors flowing through the runtime after the forward-backward pass, but the model weights themselves remain protected within the enclave boundary. This architecture, encrypting model distribution while exposing gradient computation pipelines, is standard in production FL deployments (e.g., Google's FL infrastructure).

Moreover, current state-of-the-art GIA attacks (GGL, GIFD, and DGGI) also do not actually exploit parameters. optimization-based GIA reconstructs training data by repeatedly querying the gradients of intermediate reconstructions, and the gradients from a large number of intermediates carry far more information than the parameters do.

Answer to Weakness 2.

Table 1: Defense Performance against Multi-Query Averaging:

AttackσDefenseMSE ↑
DLG (CIFAR-10)0.02ADAN0.1698 ± 0.0178
DLG (CIFAR-10)0.04ADAN0.1912 ± 0.0134
DLG (CIFAR-10)0.06ADAN0.1801 ± 0.0216
DLG (CIFAR-10)0.08ADAN0.1836 ± 0.0119
GGL (ImageNet)0.05ADAN0.0709 ± 0.0099
GGL (ImageNet)0.10ADAN0.0837 ± 0.0067
GGL (ImageNet)0.15ADAN0.0826 ± 0.0091
GGL (ImageNet)0.20ADAN0.0891 ± 0.0049

Analysis: Although the numerical metrics may vary within a certain range, the reconstructions on CIFAR-10 are essentially noise, rendering such numerical fluctuations inconsequential. For GGL, the main evaluation criterion is semantic discrepancy (i.e., visual difference).

Reviewer_EFYDOpenReview ↗

Follow up

Dear Authors,

Thank you for the follow up. I still believe a white-box evaluation is important and the reason given is unsatisfactory to me. Therefore, I will keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.