⚡ electrical engineering

A Nonlinear Incremental Approach for Replay Attack Detection

This paper proposes a nonlinear incremental approach for replay attack detection in observer-based output feedback controlled systems, introducing a watermark-based design framework that quantifies the trade-off between detection performance and control loss through incremental gains and enables the co-design of the watermark, controller, and observer.

Original authors: Tao Chen, Andreu Cecilia, Lei Wang, Daniele Astolfi, Zhitao Liu

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tao Chen, Andreu Cecilia, Lei Wang, Daniele Astolfi, Zhitao Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the security guard for a high-tech factory (a Cyber-Physical System). This factory has a brain (a computer) that talks to the machines (the physical plant) over a wireless network. The machines send back reports on how they are doing (sensor data), and the brain sends back instructions on what to do next.

The Problem: The "Ghost" Attack

One day, a hacker decides to play a prank. Instead of breaking in and smashing things, they just record the factory's normal operations for a while. Then, they hit "Play" on the recording and send those old, normal reports back to the brain, while secretly sending malicious commands to the machines to break them.

This is called a Replay Attack. It's like a thief recording your voice saying "Hello" and playing it back to your smart home security system to trick it into thinking you are home, while they actually break in.

The Old Guard's Failure:
The factory uses a standard security check called an "Innovation Detector." It basically asks: "Does the report coming in match what I expected to happen?"

The Flaw: If the factory is running smoothly and the hacker plays back a perfect recording of the past, the reports do match what the brain expects. The detector thinks, "Everything looks normal," and stays silent. The attack succeeds.

The New Solution: The "Watermark"

To catch the hacker, the authors propose adding a Watermark.

Think of this like a security guard who, every few seconds, whispers a secret, random word into the machine's ear (adding a tiny, random signal to the control input).

Normal Operation: The machine hears the secret word, does something slightly different because of it, and the sensor reports back that change. The guard checks the report, sees the change, and says, "Good, the secret word worked."
Under Attack: The hacker is playing back an old recording. That recording does not contain the new secret word the guard just whispered. When the machine tries to react to the old recording, it doesn't move the way it should. The sensor report looks "stale" or "out of sync."
The Result: The guard sees the mismatch, realizes, "Hey, this report is from yesterday! We are under attack!" and sounds the alarm.

The Catch: The "Noise" Problem

Here is the tricky part: Adding a secret word (watermark) is like adding a tiny bit of static noise to a radio station.

If you add too much noise, the music (the factory's performance) sounds terrible.
If you add too little, the hacker might not notice the difference, and the attack goes undetected.

For linear systems (simple, predictable machines), engineers already knew how to find the perfect balance. But for nonlinear systems (complex, unpredictable machines like robots with joints or chemical plants), it's much harder. You can't just use a simple formula; the math gets messy.

The Paper's Big Idea: "Incremental Gains"

The authors developed a new mathematical toolkit called Incremental Gains.

The Analogy: Imagine you are pushing a heavy box.
- Linear: If you push it twice as hard, it moves twice as fast. Easy to predict.
- Nonlinear: If you push it twice as hard, it might get stuck, slide sideways, or move three times as fast depending on the floor friction. It's unpredictable.
The Solution: Instead of trying to predict the exact path of the box, the authors developed a way to measure the maximum and minimum amount the box could move given a push. They call these "bounds."
- Lower Bound (The Detective): They ensure the watermark is strong enough that the "mismatch" during an attack is guaranteed to be big enough to be seen.
- Upper Bound (The Performance): They ensure the watermark is weak enough that the "extra movement" during normal operation stays within safe limits.

The Master Plan: Co-Design

The paper doesn't just suggest adding a watermark; it suggests redesigning the whole team to work together.

The Controller: The brain that gives orders.
The Observer: The brain that guesses what the machine is doing.
The Watermark Generator: The one adding the secret words.

Usually, these are designed separately. The authors created an algorithm (a step-by-step recipe) that designs all three at the same time. It's like training a soccer team where the coach, the players, and the strategy are all optimized together to win the game, rather than just picking random players and hoping they work together.

The Results

They tested this on a simulated robot arm (a single-link robot).

Without the new method: The robot couldn't tell the difference between a real attack and normal operation.
With the new method: The robot detected the attack almost immediately.
The Bonus: Even though they added the "noise" (watermark) to catch the hacker, the robot still moved smoothly and didn't break.

Summary

This paper solves a security puzzle for complex machines. It proves that old security guards (detectors) can be fooled by replay attacks. It introduces a new strategy (watermarking) that acts like a secret handshake. Most importantly, it provides a mathematical way to tune this handshake so it's loud enough to catch a thief but quiet enough not to disturb the machine's daily work, even when the machine behaves in complex, unpredictable ways.

1. Problem Statement

The paper addresses the critical security challenge of replay attacks in Cyber-Physical Systems (CPSs), specifically focusing on nonlinear plants.

The Threat: In a replay attack, an adversary records legitimate sensor measurements and injects them back into the system while simultaneously sending malicious control signals. This is stealthy and model-free, making it difficult to detect.
The Limitation of Existing Methods:
- Innovation-based detectors (e.g., $\chi^2$ detectors): These rely on the discrepancy between predicted and actual sensor outputs. The authors demonstrate that for nonlinear systems under certain stability conditions, these detectors may fail to detect replay attacks because the monitoring signal converges to values indistinguishable from the "no-attack" scenario.
- Linear vs. Nonlinear: While watermark-based detection (adding a random signal to the control input) is well-established for linear systems, extending these methods to nonlinear systems is difficult. Traditional linear methods rely on calculating exact covariance differences and Linear-Quadratic-Gaussian (LQG) costs, which are analytically intractable for general nonlinear systems.

2. Methodology

The authors propose a nonlinear incremental approach utilizing incremental gains to design a watermark-based detection framework.

A. Theoretical Foundation: Incremental Gains

Instead of relying on statistical distributions (which are hard to compute for nonlinear systems), the authors use incremental $L_2$ gains:

$L^+_{\delta 2}$ Gain (Upper Bound): Quantifies the maximum degradation in control performance (output difference) caused by the watermark signal.
$L^-_{\delta 2}$ Gain (Lower Bound): Quantifies the minimum difference between the system's output under attack versus the nominal case. This ensures that the watermark creates a detectable "fingerprint" that cannot be replayed.
Lyapunov and LMI Characterizations: The paper derives Lyapunov-based conditions and Linear Matrix Inequalities (LMIs) to compute these gains for nonlinear systems with linear outputs.

B. Detection Framework

Watermark Injection: A watermark signal $v = G\xi(t)$ is added to the control input. $\xi(t)$ is a non-periodic signal (e.g., chaotic or Bernoulli) to ensure it does not match historical data.
Performance Evaluation:
- Detection Performance: Guaranteed if the $L^-_{\delta 2}$ gain is sufficiently large, ensuring the monitoring signal $g(t)$ exceeds a threshold during an attack.
- Control Performance Loss: Bounded by the $L^+_{\delta 2}$ gain, ensuring the watermark does not destabilize the system or degrade performance beyond an acceptable limit ( $\alpha$ ).
Optimization Problem: The authors formulate an optimization problem to maximize the detection gain ( $\beta$ ) while constraining the performance loss ( $\alpha$ ).
- Co-design: The framework is extended to co-design the Watermark Gain ( $G$ ), the Controller ( $K$ ), and the Observer ( $L$ ) simultaneously to achieve the best trade-off.

C. Solution Algorithm

Since the resulting conditions involve bilinear matrix inequalities (BMIs), the authors propose an Iterative LMI algorithm:

The problem is relaxed into a sequence of LMIs.
The algorithm iteratively updates the variables ( $G, K, L$ ) to converge to a local optimum that satisfies the detection and performance constraints.

3. Key Contributions

Failure Analysis of Innovation Detectors: Rigorous proof that innovation-based detectors alone are insufficient for replay attack detection in nonlinear systems under specific stability conditions.
New Watermark Framework for Nonlinear Systems: Introduction of a design framework based on incremental gains ( $L^+_{\delta 2}$ and $L^-_{\delta 2}$ ), avoiding the need for explicit probability distribution calculations required in linear methods.
Extension of $L_2$ Gain Theory: Development of the incremental $L^-_{\delta 2}$ gain concept, including its Lyapunov characterization and LMI formulation, which had not been well-developed in the literature.
Co-design Optimization: A systematic approach to jointly optimize the watermark, controller, and observer gains to balance security and control performance.
Generalization: The method is applicable to a broader class of nonlinear systems compared to existing linear-specific watermarking techniques.

4. Results

The proposed framework was validated using a single-link robot system (a nonlinear plant) with two types of watermarks: Chaotic and Bernoulli.

Detection Performance:
- Without watermarks, the innovation-based detector failed to trigger alarms during replay attacks.
- With the proposed watermark, the monitoring signal $g(t)$ significantly exceeded the threshold, successfully detecting the attack.
- Co-design Impact: The co-designed parameters ( $K_{opt}, L_{opt}, G_{opt}$ ) yielded significantly better detection rates and lower detection delays compared to initial parameters.
Statistical Metrics (1000 simulations):
- Chaotic Watermark: Detection rate improved from 87.4% to 97.3% (with tighter performance constraints) and reached 100% with relaxed constraints. Average detection delay reduced from 10.28s to 8.34s.
- Bernoulli Watermark: Detection rate improved from 69.5% to 99.6%. Average detection delay reduced from 13.15s to 6.01s.
Robustness: The system maintained stability and bounded performance loss ( $\alpha$ ) across various noise levels.

5. Significance

Bridging the Gap: This work is significant because it moves replay attack detection from the well-studied linear domain into the complex, real-world domain of nonlinear systems.
Computational Tractability: By using incremental gains and LMIs, the authors provide a computationally feasible method to design secure controllers for nonlinear plants, which was previously a major theoretical hurdle.
Practical Applicability: The ability to co-design the watermark with the controller and observer allows for a more efficient trade-off between security and system performance, making the solution viable for real-world CPS applications like robotics and industrial automation.
Theoretical Advancement: The formalization of the $L^-_{\delta 2}$ gain provides a new tool for analyzing the lower bounds of system responses, which is crucial for guaranteeing detectability in nonlinear control theory.