Abstract
Video streaming is critical in IoT systems, enabling a variety of applications such as traffic monitoring and health caring. Traditional adaptive bitrate streaming (ABR) algorithms mainly focus on improving Internet video streaming quality where network conditions are relatively stable. These approaches, however, suffer from performance degradation at the IoT edge. In IoT systems, wireless channels are prone to interference and malicious attacks, which significantly impacts Quality-of-Experience (QoE) for video streaming applications. In this paper, we propose a secure and network-state-aware solution, SASA, to address these challenges. We first study the buffer-level constraint when increasing bitrate and analyze the impact of throughput overestimation in bitrate decisions. Based on these results, SASA is designed with both an offline and an online phase. In the offline phase, SASA precomputes optimal ABR configurations under various network conditions. In the online phase, SASA adopts an online Bayesian changepoint detection method to detect network changes and apply precomputed configurations to make bitrate decisions. We evaluate SASA using 429 real network traces, demonstrating that it outperforms state-of-the-art ABR algorithms such as RobustMPC and Oboe in IoT environments.
Introduction
Video streaming has become a cornerstone of modern IoT infrastructure, supporting critical services from remote health monitoring to industrial quality control. However, the "IoT edge" presents a uniquely hostile environment for data transmission. Unlike the relatively stable fiber links of home or office networks, IoT wireless links are characterized by extreme volatility, interference, and susceptibility to malicious attacks.
Traditional Adaptive Bitrate (ABR) algorithms, which determine the quality of the next video segment to download, often fail in these settings. They typically rely on moving averages to predict future bandwidth, a strategy that leads to a "throughput overestimation trap" when network conditions drop suddenly. This results in frequent rebuffering, which is particularly damaging for real-time IoT analytics that require a steady, uninterrupted video feed.
The Challenge: Throughput Overestimation
A primary cause of performance degradation in existing ABR solutions like RobustMPC is the lag in throughput prediction. When a network state shifts from high to low bandwidth, algorithms using historical averages continue to predict high throughput for several cycles.
As shown in Figure 1, the prediction (blue triangles) lags significantly behind the actual truth (red circles) at time $t_1$. This overestimation causes the player to request high-quality chunks it can no longer afford to download quickly, leading to a rapid depletion of the buffer level and eventual rebuffering events (seen at $t_3$ and $t_4$).
The SASA Framework
To address these challenges, we propose SASA (Secure and network-State-Aware), a two-stage framework designed for the volatile IoT edge.
Security and Encryption
To protect sensitive IoT data, SASA implements an RC4 stream cipher encryption at the edge server, ensuring data integrity and privacy without significant computational overhead.
Offline Phase: Optimal Configuration Precomputation
SASA utilizes a "Virtual Player" to simulate thousands of network conditions, varying both mean throughput and standard deviation. The system iterates through different "discount factors" ($d$)—safety margins that intentionally underestimate throughput—to identify which settings maximize QoE for every possible scenario. These optimal configurations are stored in a lookup table for rapid retrieval.
Online Phase: Bayesian State Awareness
Instead of simple averaging, SASA employs Bayesian changepoint detection to identify exactly when a network state has shifted. When a "sudden drop" is detected, SASA immediately pulls the optimal discount factor from its offline table. This allows the system to make conservative, stable bitrate choices for the subsequent chunks, preventing the buffer from crashing during the transition period.
Experimental Evaluation
We evaluated SASA using 429 real-world network traces (WiFi and 3G/4G) collected over three months. These traces were emulated using the Mahamahi tool to replicate the specific behaviors of IoT environments. SASA was compared against two leading ABR algorithms: RobustMPC and Oboe.
Key Results
- QoE Improvement: SASA achieved a median QoE improvement of 4.5% over RobustMPC and 4.2% over Oboe.
- Rebuffering Reduction: In over 42% of test sessions, SASA reduced rebuffering time by an average of nearly 1 second compared to RobustMPC.
- Stability: While maintaining competitive average bitrates (~2455 Kbps), SASA demonstrated significantly higher stability, avoiding the "jittery" quality changes that frustrate human viewers and disrupt automated video analytics.
Conclusion
SASA addresses the critical need for network-state-aware video streaming at the IoT edge. By combining statistical changepoint detection with precomputed optimal safety margins, it successfully avoids the throughput overestimation trap. Future work will focus on optimizing the "start-up" phase of video playback and reducing the computational cost of the offline precomputation phase, which currently requires significant time to populate the configuration table.
Summary updated