Denial-of-Service or Fine-Grained Control: Towards Flexible Model Poisoning Attacks on Federated Learning

1 Introduction

Federated learning (FL) [McMahan and Ramage, 2017] has recently emerged as a new distributed learning paradigm. It enables multiple clients (e.g., mobile devices) to jointly learn a global model without revealing their privacy-sensitive data. Specifically, in FL, numerous clients can train local models on their private datasets and iteratively upload models/updates to the central server (e.g., Apple, Google). The server will aggregate their updates to obtain global model updates using an aggregation algorithm (AGR).

However, Due to its decentralized nature, FL is susceptible to poisoning attacks. By poisoning local data, malicious clients (i.e., clients compromised by an adversary) can drive the global model to learn the wrong knowledge or embed backdoors (known as data poisoning attack) [Tolpegin et al., 2020; Wang et al., 2020]. Or they can submit malicious models to the central server to disrupt the training process (a.k.a. model poisoning attack) [Zhou et al., 2021; Xie et al., 2020], such that the learnt model suffers from high testing error indiscriminately and eventually causes a denialof-service (DoS).

These attacks can be either untargeted, with the goal of reducing the overall quality of the learnt model to cause DoS, or targeted, controlling the model to misclassify samples into an adversary's desired class (known as backdoor attack). Our work focuses on the untargeted model poisoning attack (MPA), but the outcome is not merely causing DoS. It opens up a new attack surface (dubbed fine-grained control attacks) in FL that is more stealthy than DoS.

To mitigate MPAs in FL, a growing number of defense options have been proposed [Cao et al., 2021; Prakash and Avestimehr, 2020], where the server employs a robust AGR. Their main approach is to compare the clients' local updates and exclude abnormal ones before aggregation. Recent studies [Fang et al., 2020; Shejwalkar and Houmansadr, 2021] show further adaptive or optimized MPAs can still bypass robust AGRs, especially when AGRs are known to the adversary. These novel MPAs have a huge impact on FL compared to traditional data poisoning attacks.

However, most existing advanced MPAs suffer from (part of) the following limitations:

• Assumptions about the adversary are often strong. For instance, the adversary knows the AGRs adopted by the central server [Fang et al., 2020], or can assess all benign clients' local model updates (or their compele statistics) [Baruch et al., 2019]. Practically, without these assumptions, the attack would be greatly weakened.
• Advanced MPAs [Shejwalkar and Houmansadr, 2021] commonly craft poisoned updates via maximizing the distance between poisoned and benign updates towards a prescribed direction (e.g., inverse of the global model). However, the prescribed direction itself does not guarantee good attack performance, which in turn makes the advanced MPAs built upon it less optimal.
• The adversary cannot precisely control the impact of the submitted poisoned model updates. To cause DoS, the adversary needs to craft poisoned updates farthest from the global model, which makes the attack easier to detect.

To address aforementioned challenges, we introduce a novel (untargeted) Flexible Model Poisoning Attack (FMPA). FMPA works efficiently without additional information about the FL system, except for the global model that all clients receive during normal FL training. The high-level approach of FMPA is to find the current benign reference model by exploiting the key information from the historical global models. In contrast to optimizing for the best perturbation magnitude along a prescribed malicious direction (e.g., inverse of the global model, inverse standard deviation) suggested in existing works [Fang et al., 2020; Shejwalkar and Houmansadr, 2021], FMPA fine-tunes the reference model to directly obtain a malicious model with low accuracy and small perturbations. Fig. 1 intuitively illustrates the difference between FMPA and existing attacks through visualizing the loss landscape.

Our contributions are highlighted as follows:

• We introduce a new MPA for FL risk assessment. Local poisoning models are designed by attacking vulnerable parts of neural networks and minimizing the degree of
• modification to evade detection of Byzantine-robust AGRs. Extensive experiments reveal that, FMPA achieves better attack impact, albeit with weaker and more realistic assumptions on the adversary.
• Historical data in FL has been recently employed for detecting anomalous updates [Zhang et al., 2022]. To the best of our knowledge, our FMPA is the first MPA to focus on both the vertical and horizontal information of FL systems and exploit them to assist poisoning attacks.
• We introduce a new concept of model poisoning to the community, which is not to merely target for degradation of the global model's performance indiscriminately (a.k.a. DoS), but to conceal malicious clients under a precise control of the extent of the attack. This poses an even greater threat to commercialized FL applications since an adversary can inject this subtle attack into peer FL systems to gain an advantage in competing FL services without getting noticed.

2 Background and RelatedWork

2.1 Federated Learning

Assume there is a general FL system, comprising of N clients and a central server. Each client i ∈ [N] has a local training data set Di. We define the vector g ∈ Rd as the parameters of the co-learned model, while j(g; c) denotes the loss function (e.g., cross-entropy) of g on input-output pairs c. Formally, the optimization objective of client i is Ji(g) = Ec∼Di [j(g; c)]. FL aims to find a model g⋆ that minimizes the sum of losses among clients:

where client i is weighted by βt i > 0 in the t-th iteration. Specifically, FL executes the following three steps:

• Step I: In the t-th iteration, the global model gt is synchronized by the central server to all clients or a subset of them.

• Step II: After receiving gt , the i-th client trains a new local model wt i over Di by solving the optimization problem arg minwt i Ji(wt i ) and then send its local model update ∇ti = wt i - gt to the server.
• Step III: The central server aggregates all the local model updates based on an aggregation algorithm FAGR(·) (e.g., FedAvg [McMahan et al., 2017]) to obtain global update for this round as ∇t = FAGR(∇ti∈[N] ).

Finally, the server updates the global model, i.e., gt+1 = gt - η · ∇t , where η is the global learning rate.

2.2 Poisoning Attacks on FL

Poisoning attacks are divided into two categories, namely model poisoning and data poisoning attack. This paper focuses on the more dangerous MPAs.

Model Poisoning Attack (MPA)

MPA is more dangerous as the adversary can directly manipulate local model updates. Previous MPAs include sign flipping attacks [Li et al., 2019], Gaussian attacks [Xie et al., 2018], etc. But these attacks cause a significant difference between poisoned and benign updates, which makes it easier to detect. As discussed below, state-of-the-art (SOTA) MPAs are either adaptive-based or optimization-based.

LIE [Baruch et al., 2019] computes the mean µ and standard deviation σ of all updates, and finds the fixed coefficient z based on the total number of clients. The final generated malicious update is a weighted sum: ˜∇ = µ + zσ. But LIE unrealistically assumes complete knowledge of all clients' updates. Under the assumption that the exact defenses are known to adversaries, [Fang et al., 2020] tailored attacks for different defenses by formulating it as an optimization problem of finding the optimized poisoned updates against known defenses. Derived from this, [Shejwalkar and Houmansadr, 2021] proposes Min-Max and Min-Sum, which minimize the maximum and the sum of distances between poisoned and all the benign updates, respectively. As reported in [Shejwalkar and Houmansadr, 2021], choosing an appropriate perturbation direction for the malicious update ˜∇ is the key to an effective MPA. Different heuristic directions, e.g., inverse standard deviation or inverse sign, are used by them, while our work directly investigates the optimal ˜∇. [Xie et al., 2020] devised a MPA known as inner production manipulation (IPM). MPAF [Cao and Gong, 2023] launches an attack based on fake clients injected into the FL system. But both IPM and MPAF are limited in breaking a few simple defenses.

2.3 Existing Byzantine-Robust AGRs

We divide the current defenses into three categories, namely statistics-based, distance-based and performance-based, depending on the principles that the server relies to detect or evade suspicious models. Distance-based defenses focus on comparing the distance between local updates to find outliers. Statistics-based defenses utilize statistical features to remove statistical outliers. Performance-based defenses depend on a validation dataset to evaluate the quality of uploaded models. But some recent performance-based defenses are screened by evaluating the performance of local models with the need for the central server to use a clean dataset [Wan et al., 2022; Cao et al., 2021]. This violates the privacy principle of FL, and similar to other Byzantine-Robust AGRs like [Karimireddy et al., 2021; Shejwalkar and Houmansadr, 2021], we do not evaluate these defenses in detail.

3 Flexible Model Poisoning Attack (FMPA)

In this section, before presenting our FMPA, the threat model is discussed.

Adversary's Capabilities. Suppose an adversary controls m of total N clients (i.e., malicious clients). We assume that the malicious clients in all adversarial settings are below 50% (i.e., (m/N) < 0.5), otherwise the adversary will easily manipulate FL. The adversary can obtain the global model broadcasted each round and manipulate the local updates on malicious clients.

Adversary's Knowledge. We assume a weaker adversary who neither knows the exact AGR used by the server nor any information from other (N -m) benign clients (e.g., models/gradients or local data). The adversary only has access to the data of the malicious clients as in normal FL training.

Adversary's Goals. Similar to existing MPAs, some attackers want to indiscriminately downgrade the accuracy of the global model at its testing time. Different from existing studies, some attackers may want to precisely control the effect of poisoning while maintaining concealment so as to gain advantages in establishing a competing FL service.

Next, we will introduce the method of designing concealed and effective malicious models in FL. We start with an analysis to show how to build the general model poisoning framework for deep classification models and then adapt it to flexibly poison FL systems to meet the above goals.

3.1 Model Poisoning ProblemFormulation

Consider a classification task with L classes and a DNN model Θ. For a clean input-output pair (xi, yi), we denote the output probability for the l-th category in a neuron network as Zl (xi,Θ), where xi ∈ X and its ground-truth label yi ∈ Y. Predictions for a test sample x are usually measured using the softmax function:

where P = [Pl]Ll=1 represents the confidence score vector of each category. Here, Pl ≥ 0, ∑l∈[L] Pl = 1 and v(x) is the logit vector of the penultimate layer of the neural network. Finally, the predicted label of the test sample (x, y) is ˆ y = argmaxl∈[1,L] Pl. Hereinafter, we follow the MPA literature to focus on image classification tasks for easy presentation.

The ultimate goal of any untargeted MPA is to add adversarial perturbations to network parameters to make the inference results deviate from the true labels of the images. This is similar to the case of adversarial attack [Szegedy et al., 2014; Zhang and Wu, 2020], with the difference that perturbations are added to the input images. Inspired by this observation, we aim to design an appropriate adversarial objective function to account for creating malicious models by gradientbased optimization. Generally, any untargeted attack can be defined as the following constrained optimization problem:

where H is a measure of the distance (e.g., ℓ0, ℓ2, or ℓ∞ distance) between the benign and malicious models, δp is a perturbation crafted by the adversary to poison the model Θ.

To address the difficulty of directly solving the above optimization problem (non-convex constraints and non-linear), we adopt the heuristic from [Carlini and Wagner, 2017] to express it in a form that is suitable for gradient-based optimization. Say Θ′ = Θ + δp, we define a new loss function L(Θ′ ) = L(Θ+ δp) such that the constraint y ̸= ˆ y = argmaxl Zl(x, Θ′ ) is satisfied if and only if L (Θ′ ) ⩽ 0. We use a variant hinge loss to characterize this requirement, i.e.,

where penalties are incurred when the entry of the groundtruth label is the largest among all labels.

Use the typical ℓ2-norm to measure the similarity between benign and poisoned models, the untargeted MPA problem can be represented by the following equation:

where λ is a scalar coefficient to control the relative importance of perturbation. The perturbation δp (and the malicious model Θ′ ) can be now solved by gradient-based optimization through backpropagation.

The above formulation of the general untargeted MPA reveals the key fact that, regardless of centralized or FL, the desired model perturbation δp (and thus a poisoned model) can be directly optimized. This is how the proposed FMPA differentiates from the existing works [Shejwalkar and Houmansadr, 2021; Baruch et al., 2019; Fang et al., 2020], where the standard deviation noise or the inverse of the global model (refer to Fig. 1 for visualization) is used while optimization is enforced to the magnitude of the prescribed direction perturbation to bypass defenses.

3.2 Optimization-BasedMalicious Model Update

Instead of poisoning a fixed model Θ as shown above, poisoning of FL is a sequential process along its entire training. To apply the new formulation to FL, the adversary first needs a reference model as a starting point to craft a poisoned update. The basic principle to be followed is that an ideal reference model should approximate the aggregation result FAGR(w{i∈[N]}) in the absence of attacks. Next, we illustrate how an attacker can obtain an appropriate reference model.

Different Reference Models. We discuss three reference models, namely historical reference model (HRM), alternative reference model (ARM) and predictive reference model (PRM). HRM is a naive approach that takes the historical global model gt of round t as the reference. ARM is usually adopted in existing MPAs [Fang et al., 2020; Shejwalkar and Houmansadr, 2021], and it actually simulates the average aggregation of all m local updated models w{i∈[m]} available to the attacker, i.e., 1m ∑i∈[m] wi. It is clear that the quality of ARM depends on the number m of compromised clients and their local datasets. PRM provides a better way of computing the reference model and does not depend on any other conditions. It predicts gt+1 from global historical information as discussed below.

We use exponential smoothing [Brown and Meyer, 1961], a well-known lightweight forecasting algorithm for time series, to build a PRM. Assume that gt is global model in the t-th round, s(1) t and s(2) t are the first and second order exponential smoothing values for round t, respectively. According to Taylor series and Exponential smoothing, we calculates the estimator ̂ gt+1 as the reference model:

justs the portion of the latest global model. We take the average of the first several true values to initialize s(1) 0 and s(2) 0 , and recursively computes s(1) t and s(2) t . Then we can easily update ̂ gt+1 using only s(1) t and s(2) t .

It is worth noting that unlike the literature, prediction of gt+1 is used for detecting malicious clients [Zhang et al., 2022], we are the first to use it to aid in designing MPA. As will be validated in the experiments, PRM achieved the most stable poison effect among the three reference models.

Indiscriminate Untargeted Attack (I-FMPA)

In this scenario, for each training round of FL, malicious clients upload crafted poisoned updates to affect the to-beaggregated central model. Note that we keep all poisoned updates the same to maximize the attack effect. Taking ̂ gt+1 as the reference model, malicious model can be crafted by solving Eq. (5) using stochastic gradient descent. Algorithm 1 gives a complete description of I-FMPA.

To cause DoS, the best attack performance that untargeted MPAs can achieve is random guessing. To maximize attack impact, for a classification model with L categories, we set the accuracy threshold τ in Algorithm 1 to be ( 1L ). First, the adversary divides the available data D on compromised clients into a training set Dtrain and a validation set Dval . In lines 5-6, the adversary initializes the malicious model by the reference model ̂ gt+1 . In lines 8-9, the adversary finds a certified radius R, which is subsequently used to tailor malicious updates to a suitable size for circumventing defenses. In lines 13-14, the adversary trains a fresh malicious model Θ′ with the optimization objective of Eq. (5). The desired accuracy threshold τ is ensured through validating on Dval in line 15. In line 21, the adversary conceals itself by projecting the update on a ball of radius δ (∥δ∥2 < R) around ∇t .

As mentioned in Section 2.3, we focus on two main categories of defenses, i.e., statistics-based and distance-based. Since models close to each other may well infer the same label for the same data point, more similar models imply better robustness (against poisoning attack) [Panda et al., 2022]. In these categories of defenses, outliers of uploaded model updates (which we describe with the certified radius R in the

Algorithm 1 Generating Malicious Model (Update)

Input: m malicious clients, global model gt ∈ Rd broadcasted at round t, accuracy threshold τ , malicious clients data D = {D1, . . . ,Dm} available to the adversary, the first and second order exponential smoothing s(1) t , s(2) t . Function: A(·): optimizer for parameter updates P(·, ·): Predict the next-round model (gt → ̂ gt+1 ) as Eq. (6) Output: malicious model Θ′ and model update ˜∇t 1: if t = 0 then 2: s(1) 0 ← g0 , s(2) 0 ← g0 ▷ initialize values 3: else 4: Split D into train set Dtrain and validation set Dval 5: Set ̂ gt+1 ← P(s(1) t , s(2) t ) ▷ construct estimator 6: Initialize malicious model Θ′ ← ̂ gt+1 , certified radius R ← 0, average update ∇t ← ̂ gt+1 - gt 7: for each malicious client i = 1 to m do 8: Set ∇ti ← NormalLocalTraining(gt , Di) 9: R = Max (∥∥∥∇ti - ∇t∥∥∥2 , R)▷ find certified radius 10: end for 11: for local epoch e = 1 to E do 12: for batch b in Dtrain do 13: Calculate L (Θ′ , Dtrain b ) based on Eq. (5) 14: Θ′ ← Θ′ - A (∇L(Θ′ , Dtrain b )) 15: if Accuracy(Θ′ , Dval ) ≤ τ then 16: Goto Line 20 ▷ early stopping 17: end if 18: end for 19: end for 20: Set ˜∇t ←Θ′ - gt ▷ calculate malicious update 21: ˜∇t = ∏∇t+{δ∈Rd :∥δ∥2<R} ( ˜∇t) ▷ project it on ball 22: Output Θ′ and ˜∇t 23: end if 24: Update s(1) t+1 and s(2) t+1 according to Eq. (6)

proof) are likely to be considered as suspected attackers and those updates will be reprocessed (e.g., eliminated, purified, corrected) before aggregation. However, our attack ensures that poisoned models are close to the clique of benign ones. As practiced in [Shejwalkar and Houmansadr, 2021], such a fundamental attack principle achieves good performance because it can seriously confuse defenses by avoiding making itself an outlier. Further, I-FMPA searches the best-influential poisoned model in the neighborhood of the benign reference model, while the search distance is upper bounded by the maximum distance between any normally trained model (available to the attacker) and the reference model. We also provide the theoretical analysis of I-FMPA.

Definition 1 (ϱ-Corrupted Poisoning Attack). Let ζ := {∇t1 , . . . , ∇tk . . . , ∇tN } be the union of local updates ∇ ∈ Rd among all clients at round t. The set of poisoned unions S(ϱ) consists of ζ ⋆ := { ˜∇t1 , . . . , ˜∇tm,∇tm+1 , . . . ,∇tk , . . . ,∇tN }, which is identical to ζ except the updates ˜∇t{i∈[m]} is a ϱcorrupted version of ∇t{i∈[m]} . That is, for any round t, we have ζ ⋆ = ζ + ϵ for certain ϵ with ∥ϵ∥2 ≤ ϱ.

Definition 2 (Certified Radius). Suppose ζ is a union, the mean of the updates among all clients is ∇t = 1N ∑i∈[1,N] ∇ti . We call R a certified radius for ζ if sup {∥∥∥∇ti - ∇t∥∥∥2 : i ∈ [N]} = R. For simplicity, we denote it as ζ |=R.

Lemma 1. For a union and it associated certified radius ζ |= R, we assume that ζ ⋆ ∈ S(ϱ) is a ϱ-corrupted version of ζ obtained by Algorithm 1. Algorithm 1 guarantees that, for any ζ ⋆ , ζ ⋆ |=R is still satisfied.

Proof. Algorithm 1 guarantees that the distance between poisoned updates ˜∇t{i∈[m]} and benign average update ∇t are upper bounded by the certified radius R by projecting the poisoned update onto a ball of radius δ (∥δ∥2 < R) around ∇t , i.e., ∥∥∥E[ ˜∇t{i∈[m]} ] - ∇t∥∥∥2 < R.

Example 1. Consider a scenario where the defense knows the number of attackers (e.g., 20%). Then, for safe aggregation, at least 20% of the elements in union ζ ⋆ will be reprocessed. Theorem 1 shows that there exist benign updates with the largest deviation (i.e., outliers) from the mean. They have a higher probability of being reprocessed by defenses compared to our poisoned updates.

Fine-Grained Control Attack (F-FMPA)

As emphasized in the Introduction, a new feature of our FMPA is to enable the adversary to precisely control the performance drop of the central aggregated model. For instance, when the FL training is finished, the aggregated model g′ (in the presence of fine-grained model poisoning) is inferior to model g (without model poisoning) with 10% drop in accuracy. Considering the fact that the FL central server knows neither the final accuracy of the converged model nor the clients' data, it is difficult, if not impossible, for the server to even realize that it is suffering fromMPA. In particular, competing FL service providers can easily gain an advantage with such precise and subtle MPAs, hence opening a new attack surface for FL.

To perform our fine-grained control attack (precisely manipulating the MPA effects), the adversary will set τ to his desired accuracy level, say τ = 80%, and call Algorithm 1 to produce a malicious model Θ′ (skip execution of lines 7-10 and 21). The remaining key issue is how to design model updates ˜∇Precise for malicious clients such that the aggregation result will be exactly Θ′ . This can be achieved analytically by looking at the server-side aggregation. Specifically, for FedAvg with and without MPA, we conclude:

where ∇ is the averaged model update of all the malicious clients without launching MPA. In view of this fact, it is easy to see the malicious model update is:

Replacing the unknown gt+1 with the predicted model ̂ gt+1 using Eq. (6), the goal of precisely controlling the MPA by attackers is achieved. In the case that ρ = N ηm is unknown, F-FMPA search for a suitable scaling factor ρ by iteratively increasing it each round and validate the effect on Dval . Experiments show that F-FMPA also generalizes well to other robust AGRs (especially those based on filtered averaging).

4 Experimental Results and Analyses

4.1 Experimental Setup

Datasets andModels. We conduct experiments on the following datasets and models, MNIST [Lecun et al., 1998] with a fully connected network (FC), EMNIST [Cohen et al., 2017] with a CNN, and CIFAR-10 with Resnet20/VGG-16, to evaluate FMPA under various settings.

Parameter and Attack Settings for FL. Unless otherwise specified, we assume 20% malicious clients in all adversarial setups. Given that poisoning FL with IID data is the hardest [Fang et al., 2020], following existing work [Shejwalkar and Houmansadr, 2021], our evaluation mainly considers the case where the data is IID distributed among clients (EMNIST dataset is set to be non-IID). We also investigate the effects of different non-IID degrees on FMPA.

Poisoning Attacks for Comparison. We adopt the following MPAs for comparison, i.e., AGRT [Fang et al., 2020], LIE [Baruch et al., 2019], IPM [Xie et al., 2020], MPAF [Cao and Gong, 2023], and Min-Max & Min-Sum [Shejwalkar and Houmansadr, 2021].

Evaluated Defenses. We consider eight classical defenses, i.e., Krum & Mkrum [Blanchard et al., 2017], Median [Yin et al., 2018], Trmean (Trimmed-mean) [Yin et al., 2018], Norm-bounding [Sun et al., 2019], Bulyan [Mhamdi et al., 2018], FABA [Xia et al., 2019], and AFA [Mu˜nozGonz´alez et al., 2019], as well as two newly proposed defenses, i.e., CC (Centered Clip) [Karimireddy et al., 2021] and DNC [Shejwalkar and Houmansadr, 2021].

MeasurementMetrics. Let accbenign represent the accuracy of the optimal global model without attacks and accdrop represent the accuracy degradation caused by attacks. We define attack impact ϕ = accdrop/accbenign ×100% as the degree of performance degradation of the global model. Obviously, for a given attack, the larger ϕ is, the better the attack effect is.

4.2 Evaluation Results

We compare I-FPMA with the previously described MPAs and the results are given in Tab. 1. The column No attack reports the global accuracy without attacks, while the remaining columns show the attack impact ϕ under various AGRs. In particular, AGRT is AGR-tailored (AGR is known to adversaries), while the others are all AGR-agnostic (AGR is unknown to adversaries). For fairness, all attacks except LIE are performed without knowledge of updates from benign clients (LIE uses statistics from all updates). From Tab. 1, the attack impact ϕ of I-FMPA is on average 2.4×, 2.7×, 3.5× and 2.8× higher than AGRT, LIE, IPM, and MPAF, respectively. Under the same assumptions, I-FMPA outperforms Min-Max/Sum in most cases, showing that our attack imposes a stronger threat for causing DoS than SOTAMPAs.

Besides DoS attacks, we evaluate how precisely F-FMPA can control the attack effect. For simplicity, we define the adversary's expected accuracy drop is ξ (usually between 0 - 10%). Tab. 2 shows the attack deviation (i.e., absolute value of the difference between expected and actual accuracy drop) under different combinations of defenses and ξ. We observe that F-FMPA generalizes well to other defenses, showing that 22 out of 27 (81.48%) combinations achieve an attack deviation of less than 1%.

4.3 Ablation Studies

Impact of the percentage of attackers. Fig. 2 shows the impact of various MPAs as the percentage of attackers changes from 5% to 20% on CIFAR-10 with Resnet20. We note that, due to the difference in design rationale, FMPA is always superior to existing attacks for most combinations of malicious client percentages, defenses, datasets, and models.

Effects of the reference model. We compared the attack effect of three reference models (i.e., HRM, ARM and PRM). The results are shown in Fig. 3. Among them, the performance of HRM is always poor. As expected, ARM does not have an advantage when there are fewer malicious clients. The proposed PRM performs consistently well, revealing it to be a cost-effective option.

Impact of the non-IID degree. We generate the non-IID MNIST using the method in [Fang et al., 2020] and evaluate the performance of MPAs under different non-IID degrees. The attack impact of all MPAs increases as non-IID degree increases, and I-FMPA performs better than SOTA MPAs.

Different FL scenarios. Fig. 4 shows that all MPAs become less effective when FL uses a random client sampling strategy, because the adversary cannot consistently corrupt the global model. But I-FMPA still performs the best with different total client counts and sample rates.

5 Conclusion

This paper introduces a novel model poisoning framework: FMPA. To overcome the inefficiency of attacks caused by the prescribed perturbation direction in existing works, our I-FMPA utilizes historical data in FL to iteratively optimize the malicious model through neuron perturbations to launch more effective DoS attacks. It outperforms the SOTA attacks, albeit with weaker assumptions. We also propose F-FMPA, opening up a new attack surface in FL MPAs. It enables malicious service providers to surpass competitors stealthily. Currently, FMPA still utilizes data from malicious clients. We leave it as future work to explore FMPA on zero-shot attacks.

Acknowledgments

This work was supported in part by National Natural Science Foundation of China under Grant No. 62032020, National Key Research and Development Program of China under Grant 2021YFB3101201, the Open Fund of Science and Technology on Parallel and Distributed Processing Laboratory (PDL) under Grant WDZC20205250114. Leo Yu Zhang is the corresponding author.