A new way to protect Android devices from malware
A framework called DW-FedADMM uses dynamic weighting to help devices coordinate during threat defense.
Researchers developed a framework called DW-FedADMM to help Android devices defend against threats. This method aims to improve how devices work together to find malware. It uses a process called federated learning (FL) to train models. This allows for collaborative training without collecting raw data from clients.
Android devices like smartphones and wearables face many malware attacks. These attacks include spyware, ransomware, and Advanced Persistent Threats (APTs). Traditional ways to find malware use centralized data collection. However, these methods raise privacy concerns and struggle to scale. They also face challenges across different types of Android devices.
Federated learning (FL) offers a way to train models collaboratively. In this process, a server sends a global model to a subset of clients. Each client improves the model using its own local data. The clients then send updates back to the server for coordination. This cycle repeats until the model stabilizes.
Using FL for security has several difficulties. Frequent synchronization uses a lot of bandwidth and resources. Devices may have different computing powers or lose their connections. Sometimes, only some devices participate in a round of training. These issues can make it hard to coordinate between the client and the server.
The DW-FedADMM framework combines two main ideas. It uses primal-dual coordination and performance-dependent client-side weighting. Primal-dual coordination uses consensus constraints and dual variables to manage the model. The dynamic weighting uses a coefficient to adjust the local state of a client. This coefficient is based on how well the client has performed previously.
In DW-FedADMM, a client looks at its own predictive performance. It compares this to the best performance seen from other clients. The client then uses a coefficient to blend its old state with the current global model. This new state is used to start the local update. This helps maintain continuity when clients participate only sometimes.
The researchers tested the method on three benchmarks. These are Drebin, Malgenome, and TUANDROMD. They tested different setups with 5, 10, and 15 clients. They also used different communication budgets of 5 and 10 rounds. Each setup was repeated ten times using different random seeds.
The results showed that DW-FedADMM reached 95.42% accuracy overall. It also achieved a 94.62% F1-score, 96.06% precision, and 99.45% AUC. On the Drebin dataset, it reached a seed-averaged accuracy of 97.64% ± 0.20%. It performed better than the FedADMM method in many configurations.
DW-FedADMM also had the lowest mean false-positive rate. It also had the smallest client update norm among six compared methods. Primal-dual methods performed much better than averaging-based baselines on Malgenome and TUANDROMD. In those cases, the baselines often stayed at the majority-class solution.
The benefit of dynamic weighting was most clear on Drebin. It was also most noticeable when using a 10-round budget. The weighting also helped reduce variation across the different random seeds. The results suggest that dynamic weighting acts as a stabilizing complement to coordination.
Primal-dual coordination uses variables to regulate the difference between local and global model states. A dual variable is a mathematical tool used to manage disagreement between different model states. Client-side weighting adjusts how much a client's local information influences the training. A consensus constraint is a rule that keeps local and global models similar.
The current study focused on IID client partitions. This means the data was distributed in a similar way across clients. This choice was made to isolate the effects of client fragmentation. The researchers say that looking at non-IID data is a natural next step.
Key facts
- DW-FedADMM reached an overall accuracy of 95.42% in tests.
- The method was tested on Drebin, Malgenome, and TUANDROMD benchmarks.
- The framework uses performance-dependent weighting to stabilize coordination.
- The study used communication budgets of 5 and 10 rounds.