With the rise of large language models (LLMs), increasing research has recognized their risk of leaking personally identifiable information (PII) under malicious attacks. Although efforts have been made to protect PII in LLMs, existing methods struggle to balance privacy protection with maintaining model utility. In this paper, inspired by studies of amnesia in cognitive science, we propose a novel approach, Proactive Privacy Amnesia (PPA), to safeguard PII in LLMs while preserving their utility. This mechanism works by actively identifying and forgetting key memories most closely associated with PII in sequences, followed by a memory implanting using suitable substitute memories to maintain the LLM’s functionality. We conduct evaluations across multiple models to protect common PII, such as phone numbers and physical addresses, against prevalent PII-targeted attacks, demonstrating the superiority of our method compared with other existing defensive techniques. The results show that our PPA method completely eliminates the risk of phone number exposure by 100% and significantly reduces the risk of physical address exposure by 9.8% – 87.6%, all while maintaining comparable model utility performance.

The rapid growth of robotic aerial vehicles (RAVs) has attracted extensive interest in numerous public and civilian applications, from flying drones to quadrotors. Security of RAV systems is posting greater challenges as RAV controller software becomes more complex and exposes a growing attack surface. Memory isolation techniques, which virtually separate the memory space and conduct hardware-based memory access control, are believed to prevent the attacker from compromising the entire system by exploiting one memory vulnerability. In this paper, we propose Ares, a new variable-level vulnerability assessment framework to explore deeper bugs from a combined cyber-physical perspective. We present a data-driven method to illustrate that, despite state-of-the-art memory isolation efforts, RAV systems are still vulnerable to physics-aware data manipulation attacks. We augment RAV control states with intermediate state variables by tracing accessible control parameters and vehicle dynamics within the same isolated memory region. With this expanded state variable space, we apply multivariate statistical analysis to investigate inter-variable quantitative data dependencies and search for vulnerable state variables. Ares utilizes a reinforcement learning-based method to show how an attacker can exploit memory bugs and parameter defects in a legitimate memory view and elaborately craft adversarial variable values to disrupt a RAV's safe operations. We demonstrate the feasibility and capability of Ares on the widely-used ArduPilot RAV framework. Our extensive empirical evaluation shows that the attacker can leverage these vulnerable state variables to achieve various RAV failures during real-time operation, and even evade existing defense solutions.

Unmanned Aerial Vehicles as a service (UAVaaS) has increased the field deployment of Robotic Aerial Vehicles (RAVs) for different services such as transportation and terrain exploration. These RAVs are controlled by firmware, which is often closed-source, developed by vendors, and flashed into the ROM. While these binary blobs enable off-the-shelf management of RAVs, end users (individuals or organizations) have no idea if the control firmware is designed and implemented correctly, and can only rely on firmware updates from vendors when any vulnerability is discovered. This paper proposes DisPatch, the first reverse engineering and patching framework for understanding and improving controller design and implementation within RAV firmware. DisPatch first decompiles binary instructions and recovers controller functions and core controller variables by combining control theory with program analysis using symbolic execution and data flow analysis. End users can then write a patch in a domain-specific language (DSL), which will be translated and injected into the binary firmware by DisPatch automatically. We have applied DisPatch to two instances of commodity firmware from3DR IRIS+ and MantisQ RAVs and demonstrated 100% and 80.7% accuracy respectively in the controller decompilation. We have also shown the ability to prevent severe controller performance degradation by patching two real-world bugs with in the firmware and without breaking other functionality. Finally, we show that DisPatch introduces less than 0.53% of space overhead and 1.48% of runtime overhead without violating the soft real-time deadlines. DisPatch provides the first step towards an RAV binary firmware reverse engineering and patching system to customize controller design and implementation.

The safe operation of robotic aerial vehicles (RAV) requires effective security protection of their controllers against cyber-physical attacks. The frequency and sophistication of past attacks against such embedded platforms highlight the need for better defense mechanisms. Existing estimation-based control monitors have tradeoffs, with lightweight linear state estimators lacking sufficient coverage, and heavier data-driven learned models facing implementation and accuracy issues on a constrained real-time RAV. We present Mini-Me, a data-driven online monitoring framework that models the program-level control state dynamics to detect runtime data-oriented attacks against RAVs. Mini-Me leverages the internal dataflow information and control variable dependencies of RAV controller functions to train a neural network-based approximate model as the lightweight replica of the original controller programs. Mini-Me runs the minimal approximate model and detects malicious control state deviation by comparing the estimated outputs with those outputs calculated by the original controller program. We demonstrate Mini-Me on a widely adopted RAV physical model as well as popular RAV virtual models based on open-source firmware, ArduPilot and PX4, and show its effectiveness in detecting five types of attack cases with an average 0.34% space overhead and 2.6% runtime overhead.