<p>Credit card fraud detection demands intelligent systems that ensure data privacy across institutions. Federated learning (FL) presents a privacy-preserving alternative to centralized methods but faces challenges such as vulnerability to gradient inversion attacks, high communication overhead, and degraded performance with heterogeneous (non-IID) data. Current methods such as differential privacy, secure aggregation, and statistical unlearning address these concerns to some extent but are typically associated with privacy-accuracy-computation trade-offs. In response to these limitations, we introduce SecureFed+, a novel federated learning framework whose core contribution is the <i>principled co-design</i> of three mutually reinforcing mechanisms that have not previously been unified: (i) a linearly adaptive gradient clipping (LAGC) mechanism whose threshold schedule <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\tau _t = \tau _0 + \alpha t\)</EquationSource> </InlineEquation> is calibrated per-client to local gradient statistics, directly closing the accuracy–stability gap under non-IID financial data; (ii) 2048-bit Paillier homomorphic encryption applied at the clipped-gradient level (not the model level), so that privacy is guaranteed before aggregation and gradient inversion becomes computationally infeasible; and (iii) a performance-weighted secure aggregation scheme that dynamically down-weights malicious or low-quality updates. Unlike prior work that treats convergence stability, privacy, and adversarial robustness as isolated concerns, SecureFed+’s co-design ensures these three properties are mutually consistent—producing emergent improvements in accuracy, convergence rate, and attack resilience that no single component achieves in isolation. We further provide formal convergence bounds (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(O(1/\sqrt{T})\)</EquationSource> </InlineEquation> rate under non-IID data), a robustness bound quantifying the adversarial influence at most <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\beta N\)</EquationSource> </InlineEquation> Byzantine clients can exert, and an IND-CPA privacy guarantee. Experiments on a real-world credit card fraud benchmark confirm SecureFed+ outperforms FedAvg, FedProx, FedEncrypt, FedAvg+DP, SCAFFOLD, FedNova, TAPFed, and FedML-HE, while remaining applicable to healthcare, IoT anomaly detection, and recommender-system settings.</p>

SecureFed+ a federated learning framework with adaptive gradient clipping and encrypted aggregation for credit card fraud detection

期刊论文
发表日期：

2026年4月13日

查看全文

T. Manonmani,
N. Umakanth,
S. N. Kumar,
S. Kannadhasan,
S. K. Hareni,
P. Shree Suha Tharishana

摘要

Credit card fraud detection demands intelligent systems that ensure data privacy across institutions. Federated learning (FL) presents a privacy-preserving alternative to centralized methods but faces challenges such as vulnerability to gradient inversion attacks, high communication overhead, and degraded performance with heterogeneous (non-IID) data. Current methods such as differential privacy, secure aggregation, and statistical unlearning address these concerns to some extent but are typically associated with privacy-accuracy-computation trade-offs. In response to these limitations, we introduce SecureFed+, a novel federated learning framework whose core contribution is the principled co-design of three mutually reinforcing mechanisms that have not previously been unified: (i) a linearly adaptive gradient clipping (LAGC) mechanism whose threshold schedule \(\tau _t = \tau _0 + \alpha t\) is calibrated per-client to local gradient statistics, directly closing the accuracy–stability gap under non-IID financial data; (ii) 2048-bit Paillier homomorphic encryption applied at the clipped-gradient level (not the model level), so that privacy is guaranteed before aggregation and gradient inversion becomes computationally infeasible; and (iii) a performance-weighted secure aggregation scheme that dynamically down-weights malicious or low-quality updates. Unlike prior work that treats convergence stability, privacy, and adversarial robustness as isolated concerns, SecureFed+’s co-design ensures these three properties are mutually consistent—producing emergent improvements in accuracy, convergence rate, and attack resilience that no single component achieves in isolation. We further provide formal convergence bounds ( \(O(1/\sqrt{T})\) rate under non-IID data), a robustness bound quantifying the adversarial influence at most \(\beta N\) Byzantine clients can exert, and an IND-CPA privacy guarantee. Experiments on a real-world credit card fraud benchmark confirm SecureFed+ outperforms FedAvg, FedProx, FedEncrypt, FedAvg+DP, SCAFFOLD, FedNova, TAPFed, and FedML-HE, while remaining applicable to healthcare, IoT anomaly detection, and recommender-system settings.