FedCCCU: Fairness-Aware Federated Unlearning

• The paper introduces FedCCCU, a fairness-aware federated unlearning approach that isolates unlearning impact to only the requesting client using cross-client neuron dominance and local model editing.
• It employs a constrained local editing procedure to identify and neutralize class-associated neurons based on sensitivity scores, ensuring minimal collateral performance loss for uninvolved clients.
• Empirical benchmarks under realistic non-IID conditions demonstrate that FedCCCU effectively reduces retraining overhead while maintaining accuracy and fairness across clients.

to=arxiv_search.search 大发快三是什么query":"Federated Cross-Client-Constrains Unlearning FedCCCU fairness data discrepancy arXiv","max_results":5} code 200 to=arxiv_search.search ＿老司机_code 200 大发彩票快三json [{"arxiv_id":"(Huang et al., 8 Oct 2025)","title":"Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy","authors":["Sijia Xie","Qinghua Mao","Xinghua Xie","Yanbo Wei","Yisong Xiao","Bo Bai"],"abstract":"Machine unlearning is critical for enforcing data deletion rights like the \"right to be forgotten.\" As a decentralized paradigm, Federated Learning (FL) also requires unlearning, but realistic implementations face two major challenges. First, fairness in Federated Unlearning (FU) is often overlooked. Exact unlearning methods typically force all clients into costly retraining, even those uninvolved. Approximate approaches, using gradient ascent or distillation, make coarse interventions that can unfairly degrade performance for clients with only retained data. Second, most FU evaluations rely on synthetic data assumptions (IID/non-IID) that ignore real-world heterogeneity. These unrealistic benchmarks obscure the true impact of unlearning and limit the applicability of current methods. We first conduct a comprehensive benchmark of existing FU methods under realistic data heterogeneity and fairness conditions. We then propose a novel, fairness-aware FU approach, Federated Cross-Client-Constrains Unlearning (FedCCCU), to explicitly address both challenges. FedCCCU offers a practical and scalable solution for real-world FU. Experimental results show that existing methods perform poorly in realistic settings, while our approach consistently outperforms them."},{"arxiv_id":"(Shao et al., 2024)","title":"Federated Unlearning: a Perspective of Stability and Fairness","authors":["Qi Lei","Zhenming Liu","Puyang Yang","Zhenzhen Li","Binghui Wang","Yuejie Chi"],"abstract":"This paper explores the multifaceted consequences of federated unlearning (FU) with data heterogeneity. We introduce key metrics for FU assessment, concentrating on verification, global stability, and local fairness, and investigate the inherent trade-offs. Furthermore, we formulate the unlearning process with data heterogeneity through an optimization framework. Our key contribution lies in a comprehensive theoretical analysis of the trade-offs in FU and provides insights into data heterogeneity's impacts on FU. Leveraging these insights, we propose FU mechanisms to manage the trade-offs, guiding further development for FU mechanisms. We empirically validate that our FU mechanisms effectively balance trade-offs, confirming insights derived from our theoretical analysis."},{"arxiv_id":"(Liu et al., 15 Jun 2026)","title":"pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning","authors":["Yuanjie Ma","Zihao Wang","Peigang Li","Tianbao Yang"],"abstract":"Federated unlearning (FU) enables the removal of specific data contributions from federated learning (FL) models to comply with regulations such as the General Data Protection Regulation (GDPR). However, most existing FU methods are designed for the FedAvg paradigm, where all clients share a single global model. In practice, personalized federated learning (pFL) methods such as FedPer, FedRep, Ditto, and FedBN have become widely adopted due to their superior handling of non-IID data. These methods decompose the model into shared global layers and client-specific personalized layers, fundamentally altering the semantics of unlearning, yet this setting has received little attention. We formalize FU under the pFL paradigm, identifying a tension between unlearning completeness on shared layers and personalization preservation for remaining clients. We then propose pFedUL, a layer-aware selective unlearning framework comprising three components: (1) gradient-based layer-wise contribution attribution that separately quantifies the target client's influence on shared and personalized parameters, (2) adaptive selective unlearning that applies differentiated forgetting strategies across layer types, and (3) a lightweight recalibration protocol enabling remaining clients to restore personalization with minimal overhead. We further introduce two new metrics, Personalization Preservation Score (PPS) and Cross-client Fairness Index (CFI), to evaluate pFL-specific unlearning quality. Experiments on CIFAR-10, CIFAR-100, and FEMNIST under varying non-IID settings indicate that pFedUL achieves unlearning effectiveness comparable to full retraining while maintaining an average of 97.3\% personalized accuracy for remaining clients. Compared with six state-of-the-art FU methods adapted to the pFL setting, pFedUL consistently achieves superior personalization preservation."},{"arxiv_id":"(Deng et al., 2024)","title":"Enable the Right to be Forgotten with Federated Client Unlearning in Medical Imaging","authors":["Zicheng Deng","Xiaotian Gao","Wei Wang","Ya Zhang","Yanfeng Wang","Peng Wang"],"abstract":"The right to be forgotten, as stated in most data regulations, poses an underexplored challenge in federated learning (FL), leading to the development of federated unlearning (FU). However, current FU approaches often face trade-offs between efficiency, model performance, forgetting efficacy, and privacy preservation. In this paper, we explore the paradigm of Federated Client Unlearning (FCU) to guarantee a client the right to erase the contribution or the influence, introducing the first FU framework in medical imaging. In the unlearning process of a client, the proposed model-contrastive unlearning marks a pioneering step towards feature-level unlearning, and frequency-guided memory preservation ensures smooth forgetting of local knowledge while maintaining the generalizability of the trained global model, thus avoiding performance compromises and guaranteeing rapid post-training. We evaluated our FCU framework on two public medical image datasets, including Intracranial hemorrhage diagnosis and skin lesion diagnosis, demonstrating that our framework outperformed other state-of-the-art FU frameworks, with an expected speed-up of 10-15 times compared with retraining from scratch. The code and the organized datasets can be found at: https://github.com/dzp2095/FCU."},{"arxiv_id":"([2401.15917](/papers/2401.15917))","title":"Blockchain-enabled Trustworthy Federated Unlearning","authors":["Yanzhao Wang","Yicong Zhou","Dongxiao Yu","Xiaolei Wang","Kui Ren"],"abstract":"Federated unlearning is a promising paradigm for protecting the data ownership of distributed clients. It allows central servers to remove historical data effects within the machine learning model as well as address the \"right to be forgotten\" issue in federated learning. However, existing works require central servers to retain the historical model parameters from distributed clients, such that allows the central server to utilize these parameters for further training even, after the clients exit the training process. To address this issue, this paper proposes a new blockchain-enabled trustworthy federated unlearning framework. We first design a proof of federated unlearning protocol, which utilizes the Chameleon hash function to verify data removal and eliminate the data contributions stored in other clients' models. Then, an adaptive contribution-based retraining mechanism is developed to reduce the computational overhead and significantly improve the training efficiency. Extensive experiments demonstrate that the proposed framework can achieve a better data removal effect than the state-of-the-art frameworks, marking a significant stride towards trustworthy federated unlearning."} to=arxiv_search.search ಪ್ರಜಾವjson code 200 {"query":"federated unlearning survey right to be forgotten cross-client fairness arXiv","max_results":5} to=arxiv_search.search to=arxiv_search.search ՞նչquery":"A Survey on Federated Unlearning Challenges Methods and Future Directions arXiv","max_results":3} code 200 to=arxiv_search.search 天天彩 code 200 üpjൺjson [{"arxiv_id":"(Liu et al., 2023)","title":"A Survey on Federated Unlearning: Challenges, Methods, and Future Directions","authors":["Yuyang Liu","Yajing Wang","Zhongjie Ba","Yinghao Zhao","Mingli Ding","Qinggang Meng","Yue Zhang","Wenbo Zhao","Fengyuan Xu","Kui Ren"],"abstract":"In recent years, the notion of ``the right to be forgotten\" (RTBF) has become a crucial aspect of data privacy for digital trust and AI safety, requiring the provision of mechanisms that support the removal of personal data of individuals upon their requests. Consequently, machine unlearning (MU) has gained considerable attention which allows an ML model to selectively eliminate identifiable information. Evolving from MU, federated unlearning (FU) has emerged to confront the challenge of data erasure within federated learning (FL) settings, which empowers the FL model to unlearn an FL client or identifiable information pertaining to the client. Nevertheless, the distinctive attributes of federated learning introduce specific challenges for FU techniques. These challenges necessitate a tailored design when developing FU algorithms. While various concepts and numerous federated unlearning schemes exist in this field, the unified workflow and tailored design of FU are not yet well understood. Therefore, this comprehensive survey explores the techniques and methodologies in FU providing an overview of fundamental concepts and principles, evaluating existing federated unlearning algorithms, and reviewing optimizations tailored to federated learning. Additionally, it discusses practical applications and assesses their limitations. Finally, it outlines promising directions for future research."}] to=arxiv_search.search ՞նչquery":"Communication Efficient and Provable Federated Unlearning arXiv FATS","max_results":3} code 200 to=arxiv_search.search 重庆时时 code 200 大发快三的json [{"arxiv_id":"(Tao et al., 2024)","title":"Communication Efficient and Provable Federated Unlearning","authors":["Ruoxuan Yu","Zhenghao Wang","Jinming Li","Suman Jana"],"abstract":"We study federated unlearning, a novel problem to eliminate the impact of specific clients or data points on the global model learned via federated learning (FL). This problem is driven by the right to be forgotten and the privacy challenges in FL. We introduce a new framework for exact federated unlearning that meets two essential criteria: communication efficiency and exact unlearning provability. To our knowledge, this is the first work to tackle both aspects coherently. We start by giving a rigorous definition of exact federated unlearning, which guarantees that the unlearned model is statistically indistinguishable from the one trained without the deleted data. We then pinpoint the key property that enables fast exact federated unlearning: total variation (TV) stability, which measures the sensitivity of the model parameters to slight changes in the dataset. Leveraging this insight, we develop a TV-stable FL algorithm called FATS, which modifies the classical FedAvg algorithm for TV stability and employs local SGD with periodic averaging to lower the communication round. We also design efficient unlearning algorithms for FATS under two settings: client-level and sample-level unlearning. We provide theoretical guarantees for our learning and unlearning algorithms, proving that they achieve exact federated unlearning with reasonable convergence rates for both the original and unlearned models. We empirically validate our framework on 6 benchmark datasets, and show its superiority over state-of-the-art methods in terms of accuracy, communication cost, computation cost, and unlearning efficacy."} Federated Cross-Client-Constrains Unlearning (FedCCCU) is a fairness-aware federated unlearning (FU) approach proposed for realistic federated learning deployments in which data heterogeneity and cross-client side effects are central rather than incidental. In the formulation introduced in "Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy" (Huang et al., 8 Oct 2025), FedCCCU is motivated by two claims: first, that fairness in FU is often overlooked because exact unlearning methods typically force all clients into costly retraining, even those uninvolved; second, that most FU evaluations rely on synthetic data assumptions that ignore real-world heterogeneity. FedCCCU therefore targets a setting in which only the requesting client should bear the forgetting intervention, while non-forgetting clients should be protected from collateral utility loss.

1. Conceptual position within federated unlearning

Federated unlearning is the process of removing the effect of a client or its specific data from the global model in a federated learning setting, while preserving privacy guarantees and maintaining the model's utility (Liu et al., 2023). In the broader FU taxonomy, FedCCCU belongs to client-wise unlearning, but it is distinguished by its explicit treatment of cross-client constraints: the forgetting operation is not defined solely by whether target knowledge disappears, but also by whether non-target clients retain their task-relevant knowledge under realistic heterogeneity (Huang et al., 8 Oct 2025).

The central critique behind FedCCCU is that conventional baselines optimize the wrong proxy for deployment. Exact unlearning via global retraining can be systemically unfair because all clients, including uninvolved ones, are forced into additional computation. Approximate methods based on gradient ascent, distillation, or coarse neuron manipulation can be model-wise unfair because they make broad interventions that degrade performance for clients with only retained data. FedCCCU is introduced to address both forms of unfairness simultaneously (Huang et al., 8 Oct 2025).

This emphasis is aligned with a broader line of FU theory that formalizes verification, global stability, and local fairness as distinct but competing objectives. In that framework, FU verification is measured by

$$V(\boldsymbol{w}) = \mathbb{E} \left[F_{-\mathcal{J}}(\boldsymbol{w})\right] - F_{-\mathcal{J}}\left(\boldsymbol{w}^{r*}\right),$$

global stability by

$$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$

and local fairness by

$$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$

The lower bounds

$$V(\boldsymbol{w}) + S(\boldsymbol{w}) \geq C_s \quad\text{and}\quad 2V(\boldsymbol{w}) + Q(\boldsymbol{w}) \geq C_q$$

formalize the fact that verification, stability, and fairness are fundamentally in tension under heterogeneity (Shao et al., 2024). FedCCCU can thus be read as a concrete mechanism design response to the fairness side of that trade-off.

2. Problem setting: fairness and realistic data discrepancy

FedCCCU is defined against a specific failure mode in existing FU evaluation. Synthetic non-IID setups based on label partitioning or Dirichlet noise are treated as insufficient proxies for production federations because they do not capture cross-domain feature shifts. The target deployment regime is described as "real-noniid": clients may come from genuinely different domains, with shared label spaces but very different feature spaces and resolutions (Huang et al., 8 Oct 2025).

The paper's motivating benchmark uses a mixed, cross-domain setup with MNIST, SVHN, and USPS for digits, and CIFAR10/100 and ImageNet for images, each assigned to separate clients with shared label spaces but very different feature spaces and resolutions (Huang et al., 8 Oct 2025). Within this setting, the paper argues that aggressive approximate unlearning can remove not only the target knowledge but also the knowledge used by non-forgetting clients to represent the same class under a different domain-specific feature basis.

Two fairness notions are therefore foregrounded. System fairness means that only clients requesting unlearning should bear the retraining cost. Model fairness means minimizing side-effects on non-forgetting clients’ knowledge, such as accuracy drops on their data (Huang et al., 8 Oct 2025). The latter is especially salient when multiple clients share labels but not visual statistics, since neurons associated with a class for one client may not be the same neurons that support the same class for another.

A common misconception is that retraining is automatically the fairest reference because it is exact. The FedCCCU formulation rejects that equation. The paper states that existing exact unlearning methods typically force all clients into costly retraining, even those uninvolved, and that approximate approaches make coarse interventions that can unfairly degrade performance for clients with only retained data (Huang et al., 8 Oct 2025). Another misconception is that synthetic IID/non-IID benchmarks are adequate for assessing FU; the benchmark behind FedCCCU argues that these unrealistic benchmarks obscure the true impact of unlearning and limit the applicability of current methods (Huang et al., 8 Oct 2025).

3. Mathematical formulation and cross-client constraint mechanism

FedCCCU begins from the standard FL objective

$$F(\mathbf{w}) = \sum_{k=1}^{K} \frac{n_k}{n} \mathbb{E}_{(x, y) \sim \mathcal{D}_k} \left[ \ell(\mathbf{w}; x, y) \right],$$

where $\mathbf{w}$ are model parameters and $n = \sum_{k=1}^{K} n_k$ (Huang et al., 8 Oct 2025). Its distinctive contribution is not a new global training objective, but a constrained local editing procedure that identifies neurons associated with the class to forget and filters them through cross-client dominance criteria before editing.

Class-associated neuron identification

For every neuron $w_l^k$ and example $x_i$, the client estimates an attribution score for class $c$: $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$0 In practice, the integral is approximated by a Riemann sum,

$$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$1

and the per-client score is aggregated as

$$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$2

Each client uploads only top-$$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$3 indices per class together with sensitivity scores, not raw data (Huang et al., 8 Oct 2025).

Neuron dominance computation

The cross-client constraint is encoded through a dominance ratio. For each sensitive neuron $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$4, let $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$5 denote its sensitivity for the forgetting client and $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$6 the highest sensitivity among all non-forgetting clients. The dominance ratio is

$$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$7

A small $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$8 indicates that the neuron is primarily important for the forgetting client and is therefore a good candidate for intervention. A large $$S(\boldsymbol{w}) = \mathbb{E}[F(\boldsymbol{w})] - F(w),$$9 indicates that the neuron is also important for some non-forgetting client and should be avoided to protect retained utility (Huang et al., 8 Oct 2025).

Lightweight model editing

After ranking candidate neurons for the target class by $$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$0 from low to high, FedCCCU selects the first $$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$1 dominant neurons and edits only those neurons, for example by setting weights to $$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$2 or other neutralization (Huang et al., 8 Oct 2025). The selective nature of this intervention is the operational meaning of the "cross-client constraints": the method does not attempt to erase all class-associated circuitry, only the subset dominated by the forgetting client.

This suggests a shift in FU granularity from client-level exclusion alone to client-conditioned structural editing. A plausible implication is that FedCCCU treats unlearning as a constrained model surgery problem rather than as either full retraining or unrestricted anti-training.

4. Operational workflow

The FedCCCU workflow consists of four stages (Huang et al., 8 Oct 2025).

First, upon an unlearning request, the server sends the model to all clients. Second, each client performs client-side sensitivity analysis for relevant classes and uploads indices and scores for top-sensitive neurons. Third, the server aggregates these reports, computes neuron dominance ratios for the target class to forget, and selects dominant neurons with low impact on other clients. Fourth, the server performs the lightweight model edit on only these neurons and distributes the updated model.

Two implementation properties are explicit. No global retraining is required, and uninvolved clients need no retraining, computation, or participation after the sensitivity analysis stage (Huang et al., 8 Oct 2025). The privacy posture is also narrow and concrete: only sensitive neuron indices and sensitivity scores are uploaded, not raw data (Huang et al., 8 Oct 2025). The method is therefore positioned as a practical and scalable solution for real-world FU rather than an exact statistical reconstruction of the retrain-from-scratch model.

This distinguishes FedCCCU from earlier FU workflows. FedEraser reconstructs an unlearned model from retained historical parameter updates and a calibration procedure, trading server storage for speed (Liu et al., 2020). The client-erasure method of Halimi et al. performs local unlearning at the client to be erased and then a few FL post-training rounds among remaining clients, without storing historical updates (Halimi et al., 2022). FCU in medical imaging likewise uses local unlearning followed by rapid post-training, but shifts forgetting to the feature level via model-contrastive unlearning and frequency-guided memory preservation (Deng et al., 2024). FedCCCU instead focuses its novelty on fairness-aware localization of the intervention under real heterogeneity (Huang et al., 8 Oct 2025).

5. Empirical behavior and comparison with baseline families

The experimental comparison in the FedCCCU study is organized around three baseline families: Delete-Retrain, Relabel-Poison, and Neuron-Zeroing (Huang et al., 8 Oct 2025). Evaluation tracks both global accuracy and per-client class accuracies, which is essential because the fairness claim concerns the distribution of damage across clients rather than aggregate utility alone.

The reported qualitative pattern is consistent across the benchmark. Aggressive zeroing causes near-total forgetting but also drops accuracy of the target class for non-forgetting clients, sometimes to near-zero, which the paper describes as extreme unfairness (Huang et al., 8 Oct 2025). Global retraining only partly forgets, with class accuracy drops of approximately $$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$3–$$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$4, and still often degrades global accuracy (Huang et al., 8 Oct 2025). FedCCCU is described as nearly as strong as Zeroing in forgetting effect while preserving non-forgotten class and client accuracy substantially better than Zeroing, Relabel, or Delete-Retrain (Huang et al., 8 Oct 2025).

The central empirical claim is therefore not merely that FedCCCU forgets, but that it balances forgetting and collateral preservation better than the compared baselines under realistic heterogeneity. The paper also states that existing methods perform poorly in realistic settings, while FedCCCU consistently outperforms them (Huang et al., 8 Oct 2025). Because the benchmark is organized around per-client effects, this performance claim is explicitly fairness-aware rather than purely global.

This evaluation philosophy resonates with later FU work that introduces client-sensitive metrics. In personalized FL, pFedUL formalizes a tension between unlearning completeness on shared layers and personalization preservation for remaining clients, and evaluates it with Personalization Preservation Score (PPS) and Cross-client Fairness Index (CFI) (Liu et al., 15 Jun 2026). Although FedCCCU predates those metrics, the alignment is conceptually direct: both lines treat retained-client damage as a first-class FU criterion rather than a secondary side effect.

6. Relation to adjacent FU methodologies

FedCCCU occupies one point in a larger design space of FU mechanisms.

A survey perspective distinguishes retraining, fine-tuning, gradient ascent, multi-task unlearning, model scrubbing, and synthetic-data-based approaches, and emphasizes FU-specific challenges such as knowledge permeation, data isolation, who-unlearn, and what-to-unlearn (Liu et al., 2023). Within that landscape, FedCCCU is best understood as a selective structural editing method for client-wise unlearning under cross-client constraints.

Other major FU lines solve different bottlenecks. Provable exact unlearning is addressed by FATS, which introduces total variation stability and exact sample-level and client-level unlearning with communication efficiency (Tao et al., 2024). Provable sequential unlearning for FedAvg is addressed by SIFU, which calibrates Gaussian noise to a sensitivity proxy and supports a sequence of unlearning requests with $$Q(\boldsymbol{w}) = \sum_{i \notin \mathcal{J}} p_i' \left| \Delta f_i(\boldsymbol{w}) - \overline{\Delta f} \right|.$$5-style guarantees (Fraboni et al., 2022). Cross-client auditability is pursued by blockchain-enabled trustworthy federated unlearning, where Chameleon hash functions support a proof of federated unlearning protocol and cross-client removal verification (Lin et al., 2024). Privacy-preserving certified removal is addressed by Starfish, which uses Two-Party Computation and shared historical client data between two non-colluding servers (Liu et al., 2024).

On the utility-preservation side, later work develops more explicit conflict handling. FedOSD calculates an orthogonal steepest descent direction to be non-conflicting with other clients' gradients and closest to the target client's gradient (Pan et al., 2024). FedCARE uses conflict-aware projected gradient ascent and relearning-resistant recovery, with support for client-, instance-, and class-level unlearning (Li et al., 30 Jan 2026). FUPareto frames unlearning as Pareto-augmented optimization and uses Null-Space Projected MGDA to support fair, concurrent unlearning for multiple clients (Wang et al., 2 Feb 2026). FUSED uses selective sparse adapters to mitigate indiscriminate unlearning of cross-client knowledge and make unlearning reversible (Zhong et al., 28 Feb 2025). These developments suggest that FedCCCU's core concern—protecting non-forgetting clients from indiscriminate interventions—became a durable theme in subsequent FU research.

7. Limitations, interpretation, and research directions

FedCCCU is explicitly motivated by fairness and realistic heterogeneity, but it does not claim to solve all FU desiderata simultaneously. The broader theory of FU under heterogeneity indicates that no single mechanism can simultaneously offer perfect verification, global stability, and local fairness under significant heterogeneity (Shao et al., 2024). FedCCCU prioritizes fairness-aware localization of the intervention; this suggests that its main contribution is to alter the geometry of the forgetting action rather than to provide exact retraining equivalence.

A second limitation is scope. The published description centers on class-associated neurons and class forgetting under cross-domain client discrepancy (Huang et al., 8 Oct 2025). A plausible implication is that extension to simultaneous multi-client removal, personalized FL architectures, or cryptographically enforced hidden unlearning would require additional machinery of the kind later explored in FUPareto, pFedUL, and EFU (Wang et al., 2 Feb 2026, Liu et al., 15 Jun 2026, Mohammadi et al., 11 Aug 2025).

A third point concerns verification. The FedCCCU description emphasizes behavioral fairness and deployment realism rather than exact or certified unlearning. In the FU literature, verification has been pursued through statistical indistinguishability, blockchain auditability, membership-inference evaluation, and secure multiparty protocols (Tao et al., 2024, Lin et al., 2024, Liu et al., 2024). This suggests a natural research direction: combining FedCCCU-style cross-client constraints with explicit certification or cryptographic enforcement.

In the broader history of FU, FedCCCU marks a transition from asking whether a client can be forgotten to asking which other clients are harmed when that forgetting is operationalized. Its distinctive claim is that real-world FU must be evaluated not only by deletion efficacy and runtime, but also by fairness under genuine data discrepancy (Huang et al., 8 Oct 2025).