Introduction
Traditional Distributed Denial of Service (DDoS) attacks typically rely on high-volume traffic to overwhelm network resources. In contrast, Low-Rate DDoS (LDoS) attacks are stealthy, exploiting application-level vulnerabilities—such as Regular Expression Denial of Service (ReDoS)—to consume significant CPU or memory resources with minimal traffic volume. These attacks often bypass standard intrusion detection systems.
In cloud environments, mitigating LDoS via Virtual Machine (VM) scaling is often inefficient. VM scaling is "coarse-grained," slow, and costly, as it scales entire application stacks regardless of which specific component is under attack. This research, published in IEEE Transactions on Parallel and Distributed Systems, investigates how the lightweight, loosely coupled nature of containerized microservices can be leveraged for more surgical and dynamic defense.
Proposed Methodology
The researchers propose a dynamic mitigation strategy based on $M/M/c$ queueing theory, modeling a microservice as a system of parallel containerized servers. The strategy involves three critical phases:
- Traffic Partitioning: Incoming requests are categorized into a "Whitelist" of trusted users and "Unknown" traffic, which contains a mix of benign and potentially malicious requests.
- Resource Isolation: The system calculates the minimum resource allocation and the optimal number of containers ($c_w$) required to guarantee the Quality of Service (QoS) for whitelisted users, ensuring their average staying time remains within acceptable bounds.
- Dynamic Optimization: For unknown traffic, the system determines an optimal number of containers ($c_u$) that maximizes the survival rate of benign requests without causing performance degradation due to management overhead.
Experimental Evaluation
To validate the model, the authors developed a custom testbed using a Docker Swarm cluster. They implemented a Python-based "products" microservice with a known ReDoS vulnerability. Benign traffic was modeled after real-world e-commerce inquiry data (approx. 100 requests/second), while LDoS attacks were simulated to target the vulnerable service. Extensive MATLAB simulations were also conducted to verify the queueing model across thousands of scenarios.
Results and Analysis
The study revealed several key insights into container performance under stress:
The "More is Less" Paradox
A critical finding is that blindly increasing the number of containers in a resource-constrained environment can be counterproductive. As the number of containers grows, the OS-level overhead for managing them increases, eventually leading to a plateau and then a decline in the overall serving rate.
Model Accuracy and QoS
The mathematical model demonstrated exceptional accuracy, with a fit of $R^2 > 0.995$ compared to experimental results. Under active LDoS attack, the partitioning strategy successfully maintained the performance for whitelisted users at levels nearly identical to non-attack conditions. Furthermore, the optimized scaling for unknown traffic significantly outperformed "worst-case" blind scaling, maintaining service availability even when malicious requests comprised 30% of the unknown traffic pool.
Conclusion
This research demonstrates that containerized microservices offer unique opportunities for defending against stealthy LDoS attacks. By employing surgical scaling and physical resource isolation (via cgroups and namespaces), cloud providers can prevent "Economic Denial of Sustainability" (EDoS) and ensure high availability for critical users. Future work will look to expand this model to include multi-dimensional resource optimization (memory, I/O) and multi-threaded container architectures.
Summary updated