Abstract
This paper exposes a critical security flaw in federated learning (FL) for embodied agents: malicious clients can "hijack" a robot's navigation via a backdoor attack called Navigation as Wish (NAW). To counter this, the authors propose Prompt-Based Aggregation (PBA), a defense mechanism that uses small, learnable "prompts" to detect and filter out malicious updates by identifying disruptions in how the agent aligns visual cues with language instructions.
Introduction
Vision-and-Language Navigation (VLN) involves training agents, such as home robots, to follow natural language instructions in complex 3D environments. Because training data—often video of private homes—is highly sensitive, researchers utilize Federated Learning (FL). In this paradigm, a central server aggregates model updates from individual "clients" (homes) without ever seeing the raw data.
However, this privacy-preserving architecture introduces a "black box" vulnerability. A malicious client can "poison" its local training data to implant a backdoor. For example, an attacker could train the agent to ignore instructions and instead head toward a specific room whenever it detects a "trigger," such as a specific sticker or noise pattern. Existing FL defenses typically look for statistical outliers in model weights, but in VLN, the highly non-identical and independently distributed (Non-IID) nature of different homes makes it nearly impossible to distinguish a unique house layout from a malicious attack.
The Navigation as Wish (NAW) Attack
The authors introduce the Navigation as Wish (NAW) attack, where an adversary modifies local training trajectories. When a specific visual trigger is present, the agent is trained to move toward a target viewpoint chosen by the attacker, regardless of the language instruction. This attack is particularly insidious because it can achieve a nearly 100% Attack Success Rate (ASR) without degrading the agent's performance on legitimate, non-triggered instructions, making it invisible during standard testing.
Prompt-Based Aggregation (PBA) Defense
The core innovation of this work is the Prompt-Based Aggregation (PBA) defense. PBA introduces small, learnable vectors—prompts—added to the visual and text features before they reach the model's attention layer.
The Insight: Alignment as a Signal
Benign clients consistently attempt to align specific words (e.g., "corridor") with corresponding visual features. Malicious clients inherently break this alignment because they train the model to prioritize the trigger over the text.
The Mechanism
- Initialization: The server sends a global prompt to all clients.
- Local Adaptation: During local training, these prompts adapt to the client's specific data.
- Detection: When updates are returned, the server calculates the similarity between prompts. Because malicious updates disrupt vision-language alignment, their prompts become detectable outliers.
- Filtering: The server uses a selection rule (similar to MultiKrum) to filter out these malicious updates before aggregation.
Experimental Evaluation
The researchers evaluated PBA on two standard VLN benchmarks: Room-to-Room (R2R) and the larger, multi-lingual Room-across-Room (RxR).
- Defense Superiority: PBA significantly outperformed state-of-the-art defenses. On the RxR dataset, it decreased the attack success rate by approximately 40% compared to existing methods.
- High Fidelity: Unlike defenses that trade off accuracy for security, PBA maintained high fidelity, ensuring the global model performed as well on legitimate tasks as a model trained in a completely benign environment.
- Efficiency: By focusing on a tiny set of prompt parameters rather than millions of model weights, PBA is computationally efficient and focuses on the most stable signal in multimodal tasks: the relationship between vision and language.
Conclusion and Future Work
PBA demonstrates that monitoring the alignment of multimodal inputs is a more robust defense than traditional weight-based analysis in Non-IID federated environments. While effective against current black-box attacks, future research is needed to address more sophisticated "white-box" scenarios where attackers might attempt to mimic benign prompts or use more subtle, adaptive triggers.
Summary updated