A New Perspective on Privacy Protection in Federated Learning with Granular-Ball Computing

Federated learning (FL) is an emerging collaborative training paradigm that enables participants to train models while preserving data privacy jointly Yang et al. (2020); Gong et al. (2020). As FL continues to evolve, the importance of privacy protection has become increasingly prominent. Effectively safeguarding participants’ data privacy has become a central research focus, leading to continuous exploration and refinement of privacy-preserving techniques in existing methods Zhang et al. (2023).

Currently, privacy protection methods in FL primarily include differential privacy Le Ny and Pappas (2013); Chen et al. (2024); Jiang et al. (2024), secure multi-party computation Bonawitz et al. (2017); Bell et al. (2020), and homomorphic encryption Aono et al. (2017); Xu et al. (2020). These methods introduce innovations from the perspectives of models and features, enhancing privacy protection capabilities. However, few methods consider privacy protection from the perspective of input. Our research is the first to address this gap and propose a methodological approach in this area.

Moreover, performance and efficiency are equally crucial in FL Zhang et al. (2022, 2023). Although, due to the "No-Free-Lunch" principle, achieving significant improvements in all three aspects—privacy, performance, and efficiency—simultaneously is challenging, striving for a balance among them is essential. While previous works have made progress in specific areas, such as FedProx Li et al. (2020a) and FedOPT Ahmed et al. (2024) mitigating performance degradation due to heterogeneity and Scaffold Karimireddy et al. (2020) and FedNova Wang et al. (2020) improving efficiency and speed, these advancements remain incomplete. Our proposed framework aims to enhance efficiency while minimizing the loss in accuracy, thereby providing models with stronger generalization capabilities.

Granular Ball Computing (GBC) is a novel approach within the domain of multi-granularity computing, widely applied in machine learning and data mining. Unlike traditional methods that rely on the finest granularity of input data, such as pixel-level inputs for neural networks, GBC aligns more closely with the way the human brain processes information. Chen’s experiments demonstrated that in image recognition, the human brain prioritizes large-scale contour information over fine-grained details Chen (1982). This suggests that traditional machine learning models, which use the finest granularity inputs, are not only more susceptible to noise but also less interpretable, and misaligned with human cognitive processes.

Wang first introduced the concept of large-scale cognitive rules into granular computing, leading to the development of multi-granularity cognitive computing Wang (2017). Building on this, Xia et al. proposed Granular Ball Computing, where different sizes of hyperspheres represent varying levels of granularity: large granular balls for coarse granularity and small balls for fine granularity Xia et al. (2020, 2022). Recent advancements have further refined this approach. Shuyin et al. (2023) introduced an enhanced granular ball method for handling non-Euclidean spaces and capturing fine-grained boundary details. By using graph nodes to represent rectangular image regions instead of individual pixels, this method significantly reduces the amount of input data while preserving essential information. This not only enhances efficiency by minimizing redundant data processing but also strengthens privacy protection.

This deeply prompts us to consider the potential of GBC in enhancing privacy protection within FL. By focusing on coarse-grained and abstract representations rather than fine-grained data points, GBC naturally reduces the exposure of sensitive information. This characteristic presents a critical advantage in privacy-sensitive contexts such as FL.

FL is a machine-learning scenario with $K$ clients (denoted as $\mathcal{C}=\{C_{1},C_{2},\ldots,C_{K}\}$), and a central server (denoted as $S$). Each client has its private dataset $\mathcal{D}_{i}$. The goal of the entire system is to minimize the sum of the loss functions on all clients:

where, $N=\sum_{k=1}^{K}N_{k}$ is the total number of samples on all clients.

A common attack in FL is the data reconstruction attack Zhu et al. (2019), which reconstructs the original training data using the model’s weights and gradients with the L-BFGS algorithm, posing a significant data security risk.

The loss function $\mathcal{L}(x,x^{\prime};\mathcal{F})=||\nabla\mathcal{F}(x)-\nabla\mathcal{F}(x^{\prime})||^{2}$ is the optimization objective for the reconstruction attack. This loss function is typically used to measure the difference between the original data $x$ and the reconstructed sample $x^{\prime}$ and can be a $L_{p}$ norm or other forms of divergence measures. We use $\mathcal{F}$ to represent the network during the attack process. Consider the impact of a small change $dx^{\prime}$ in $x^{\prime}$ on the loss function $\mathcal{L}(x^{\prime},x^{\prime}+dx^{\prime};\mathcal{F})$:

where $\Delta\mathcal{L}$ represents the change in the loss function and $J(x^{\prime})$ represents the Jacobian matrix of the loss function $\mathcal{L}$ with respect to $x^{\prime}$.

During the iteration process, $x^{\prime}$ is updated as follows:

where $H$ is an approximation of the Hessian matrix, $g$ is the gradient of the current inferred data $x^{\prime}$, and $H^{-1}$ is estimated using the changes in parameters and gradients over the past few iterations. This approximation better reflects the local curvature information of the current parameters.

We want the loss function to decrease. Thus, we need:

The L-BFGS algorithm updates parameters by estimating the inverse of the Hessian matrix. At each iteration, it approximates the inverse matrix of the Hessian at the current parameter value $x^{\prime}$. Therefore, the direction of $dx^{\prime}$ is influenced by the inverse of the Hessian matrix. As the input data decreases, the magnitude of $dx^{\prime}$ becomes smaller. This means that the influence of $dx^{\prime}$ in the algorithm weakens, and parameter updates are more influenced by the inverse of the Hessian matrix. The L-BFGS algorithm can more accurately estimate the direction of parameter updates, leading to faster convergence to the minimum of the loss function.

Figure 2 illustrates the trend of the information of data reconstructed by the model as the input information changes, which also validates the theory we proposed above. When input data decreases, the reconstruction attacker can better restore the data, but the absolute information content of the obtained data will not exceed that of the original image. Therefore, considering FL’s privacy protection from the input perspective is feasible.

We can decompose the expected generalization error of FL, analyzing the relationship between input and utility. For a given model $\mathcal{F}$ and data $\mathcal{D}$, the expected generalization error of FL can be expressed as:

In Equation 5, we decompose the generalization error into model bias, variance, and data noise. Model bias depends on the model’s fitting capability and the proportion of effective information. As the amount of ineffective data decreases, both variance and data noise decrease, while model bias remains unchanged. Conversely, when the amount of effective data decreases, variance and data noise also decrease, but model bias increases. This suggests that it is feasible to retain only the effective information from the input data and reconstruct it, as long as the model’s fitting capability is preserved, thereby maintaining the utility of FL.

From an efficiency perspective, for a fixed model and algorithm, when the information content of the input data decreases, the total computation required by the model reduces, thereby enhancing the efficiency of FL.

Our proposed framework comprises two main components: granular-ball-based knowledge reconstruction and graph input-based aggregation with variance constraints. For the knowledge reconstruction, we developed a two-dimensional binary search algorithm based on variance constraints. Figure 3 illustrates the knowledge reconstruction process of this algorithm. For the input data, we first compute the gradient map using the Sobel operator. Next, we iteratively select the point with the smallest gradient that has not yet been chosen to generate a granular-rectangle, continuing this process until all points have been processed. During the generation of each granular-rectangle, we first fix its height and determine the appropriate length using a bisection method based on purity constraints. Then, with the length fixed, we similarly determine the appropriate height. When no further granular-rectangle can be generated, we reconstruct all the granular information into a graph: each granularity becomes a node, and an edge is formed between any two granular-rectangle that share overlapping pixels. In this way, the image is reconstructed into a graph. We then apply an appropriate graph neural network for federated learning, and perform model aggregation based on variance constraints, before finally uploading the aggregated results to the server.

Equation 4 and Equation 5 inspired us to consider how to balance privacy, efficiency, and utility in FL from the input perspective. From the perspectives of utility and efficiency, the performance of a model often depends on the amount of information in the data itself rather than the form in which the data is stored. This insight led us to consider privacy protection from the standpoint of data storage forms. Simply adding noise or blurring the original images can interfere with machine classification accuracy, but reconstructed images by such methods are often still recognizable to humans.

As shown in Figure 2, with very little information, the model cannot effectively learn the features of a panda, but for humans, it is still recognizable. Therefore, we need to consider filtering and reconstructing the input data to ensure that the reconstructed images are difficult for both humans and machines to recognize. The granular computing methods discussed in Xia et al. (2019) and Shuyin et al. (2023) focus on extracting and integrating input information, which is then fed into the model at a coarser granularity. This approach aligns with our concept of reconstructing input information to achieve privacy protection in federated learning. Building on this, we have designed an algorithm for the GrBFL framework.

To achieve image segmentation from fine to coarse granularity, a straightforward approach is to group adjacent similar pixels into the same granular-ball. This allows us to retain the original image structure while removing redundant information. It is important to note that the term "granular-ball" here does not literally refer to a spherical shape, as the image itself is two-dimensional. Therefore, using a rectangle is more appropriate for modeling regions within the image. We first compute the gradient map of the image using operators such as Sobel. For each pixel, the greater the gradient, the fewer similar pixels are typically found in its vicinity. Therefore, we sort the pixels in ascending order of their gradient values and construct granular-rectangle for each pixel accordingly. Purity is a classical metric for assessing the quality of granular-ball generation, which we extend to the context of image granular-rectangles. In image processing, the purity of a granular-rectangle is defined as the ratio of normal pixels to the total number of pixels within the granular-rectangle. Here, normal pixels refer to those whose gray-scale values differ from the granule’s center pixel by no more than a specified threshold. The calculation of purity is shown in Equation 6.

where $k$ is the ID of granular-rectangle, $p_{i}$ is the centre, $c_{k}$ is the other pixel points within the granular-rectangle, $thr$ is the defined threshold, $r_{x}$ and $r_{y}$ are the length and width of the granular-rectangle, $R_{k}$ is the granular-rectangle, $f(\cdot)$ is gray-scale value.

Identifying the largest granular rectangle that satisfies the purity constraint is a critical challenge that requires careful consideration. Evaluating each pixel individually can lead to prohibitively long processing times, especially when dealing with large-scale image data. However, it is noteworthy that the purity function demonstrates global monotonicity: as the granular rectangle expands in any direction, its purity consistently decreases. By leveraging this property, a binary search strategy can be employed, starting with a larger size and progressively narrowing down, to efficiently determine the optimal dimensions of the granular rectangle that satisfy the purity constraint. This approach significantly reduces the algorithm’s complexity to a logarithmic scale.

After obtaining all the granular rectangles, we construct a graph structure by abstracting the center of each granular rectangle as a node in the graph. Additionally, we consider the pairwise overlapping relationships between granular rectangles. If two granular rectangles overlap, an edge is added between the corresponding nodes; otherwise, no edge is present between the nodes. This graph construction method preserves the original structural information while also encapsulating the relationships between granular rectangles. Formally, the reconstructed graph is defined as:

In the FL scenario, each client processes the input image data as described above, resulting in a series of graph data. The GrBFL framework then inputs these graph data into the graph neural network for training. After a certain number of training rounds, the models from each client will be aggregated. The specific process of model aggregation will be detailed in the following section.

Through extensive experiments, we found that, compared to traditional graph datasets like Cora, the datasets generated from granular-ball computing exacerbate the differences between client models in FL. This increased disparity poses challenges for methods that rely on directly weighted averaging of model parameters and gradients, such as slower convergence rates. As shown in Table 1, various FL aggregation strategies were evaluated on our dataset, with the results indicating that the simple FedProx Li et al. (2020a) aggregation strategy performed better. The proximal term introduced in FedProx effectively mitigated the instability of graph neural networks when processing granular-ball inputs. The FedProx method is an evolution of the FedAvg McMahan et al. (2017) algorithm, which reduces discrepancies in model parameter adjustments by incorporating a proximal term into the loss function of the client models. This not only accelerates the convergence of the global model but also alleviates the instability encountered by graph neural networks in the context of granular-ball inputs.

The FedProx algorithm locally introduces a proximal term to limit parameter updates, i.e. $F_{k}(\cdot)$ becomes $h_{k}(\cdot)$, thus minimizing the following loss function:

Following aggregation under the FedProx mechanism, the server transmits the graph neural network model back to the client, thereby concluding a round of FL.

The time complexity of our method for processing gradients and sorting is $O(nm\log(nm))$, where $n$ and $m$ represent the length and width of the image, respectively. When solving for the granular-rectangle size using a binary search algorithm, the time complexity is $O(k\log n\log m)$, where $k$ is the number of regions and $k\ll nm$. Constructing the granular-rectangle graph has a time complexity of $O(k^{2})$. This method is highly efficient for large-scale image processing. Testing shows that for common 224x224 image data, our method completes data reconstruction in just $1ms$, thereby ensuring model performance.

Datasets. We conducts a detailed analysis on three classic image datasets: MNIST, CIFAR-10, and CIFAR-100. The varying difficulty levels of these datasets allow for a comprehensive evaluation of the model’s overall performance.

To conduct a thorough evaluation of the model within the context of FL, which includes the aspects of privacy, efficiency, utility, and their comprehensive considerations, we carefully define the following metrics.

where $img_{true}$ refers to the actual image, $img_{pred}$ refers to the image reconstructed by the attacker, and $MSE$ is mean square error.

where, $\sigma(\cdot)$ denotes the sigmoid function, which is a crucial component in calculating communication efficiency. The variables time and traffic represent the duration of communication and the volume of data communicated, respectively. Specifically, traffic is quantified as the product of the number of model parameters and the communication frequency. Additionally, $\phi$ is a scaling factor used to uniformly map the metric values to a range between 0 and 1.

The GrBFL framework demonstrates exceptional efficacy in privacy protection.

To rigorously evaluate the privacy protection capabilities of the GrBFL and CNNFL frameworks, this study conducted reconstruction attack experiments under the GCN and ResNet models. The experiments were performed across three datasets. For each dataset, a set of 100 random training samples was systematically extracted and subjected to 300 iterations to compute $S_{p}$. The attack model for CNNFL can be referenced in Zhu et al. (2019). The attack model for GrBFL is similar to that of CNNFL. In this case, the attacker can access the weights of the graph neural network and the gradients of the data, using continual optimization to reconstruct the original graph data. We assume that the attacker understands the semantics of node features. However, since node features only contain coarse pixel information, the attacker can only determine the position and size of each granule rectangle but cannot precisely obtain the specific pixel values.

Figure 4 illustrates the results of our privacy protection experiments. The upper section compares the differential privacy performance of GrBFL, LotteryFL, and Differential Privacy (DP) across different numbers of iterations. It can be observed that, although LotteryFL and DP provide some degree of privacy protection, reconstruction of the data remains possible as the number of iterations increases. In contrast, GrBFL demonstrates significant effectiveness in privacy protection. Even with a surge in the number of iterations, GrBFL’s inherent information loss mechanism effectively hinders the attacker’s ability to accurately reconstruct the training data.

Figure 4 quantitatively displays the results of the privacy protection experiments. The line chart in the lower left quadrant shows the trajectory of mean squared error (MSE) loss between reconstructed images and the original images for different scenarios, indicating that GrBFL maintains a significant divergence in data recovery, thereby validating its effectiveness in privacy protection. The bar chart in the lower right section presents the average $S_{p}$ results for 100 random samples from three datasets, demonstrating that GrBFL exhibits strong privacy protection capabilities across all datasets.

Additionally, we conducted experiments on membership inference attacks. The detailed results of additional experiment results, along with the details of experiments mentioned earlier, can be found in the supplementary materials.