Adversarial Observability and Performance Trade-offs in Optimal Control

arXiv:2506.19829 · eess.SY, cs.SY, math.OC · Submitted 2025-06-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Adversarial Observability and Performance Trade-offs in Optimal Control".

Dev: Here's the gist of the paper. We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints<ref:2506.19829#pg2>.

Rosa: The paper quantifies the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system<ref:2506.19829#pg2>.

Dev: It derives theoretical lower bounds on these metrics under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off<ref:2506.19829#pg2>.

Rosa: Finally, it shows that the performance-constrained optimization of the Gramian’s trace can be formulated as a one-shot semidefinite program, while it addresses the optimization of its inverse through sequential semidefinite programming<ref:2506.19829#pg2>.

Paper summary: Rosa: So we're looking at this paper called Adversarial Observability and Performance Trade-offs in Optimal Control by Filippos Fotiadis and Ufuk Topcu. The main idea here is developing a feedback controller that makes it hard for an adversary to observe a set of sensors on a linear system while still keeping the control performance pretty good.

Dev: Yeah, what does that actually mean in practice? Is it just about making some signals noisy or something?

Rosa: It's more about the structure of the system and how we design the control law, like Kx(t) and Kp Rˆn, so that even if an adversary has access to these sensors, they can't easily figure out what's going on inside.

Taro: From my side, I'm thinking about what happens when the world misbehaves—how does this controller handle those tricky situations?

Dev: Exactly. The paper is quantifying this adversarial observability using two specific measurements: the trace of the observability Gramian and the trace of its inverse. This lets them capture both how observable something is on average and which directions are least observable for the system Rosa How does that help us understand what's fundamentally possible?

Rosa: They go a step further by deriving theoretical lower bounds for these metrics based on how far we are from optimal performance Dev That’s interesting because it sets the absolute limits on how much we can reduce observability depending on the trade-off we make with performance Taro So, if I want to reduce observability a little bit, what's the cost?

Dev: Well, they show that smaller performance trade-offs generally lead to a sharper decrease in that trace of the observability Gramian Rosa But larger performance losses cause a steeper increase in the trace of the Gramian's inverse Taro Which means you can get better observability reduction for less performance loss if you keep things tight Dev It’s a direct trade-off between those two metrics, and they characterize these fundamental limits as a function of that trade-off distance Rosa

Taro: That sounds like it sets up a very specific boundary for what's achievable in real control systems, right? So, the core mechanism here is about finding that sweet spot where you minimize the observability while staying within those performance constraints Dev

Rosa: Right. And they tackle this optimization problem in two ways. First, optimizing the trace of the Gramian—that’s Problem one—they show it can be solved as a one-shot semidefinite program or SDP Taro That sounds computationally efficient for real-time systems, which is important because we need fast solutions Dev <ref:2506.19829#pg1,as a one-shot semidefinite program>

Dev: True. But then for optimizing that inverse, Problem two they address it using sequential semidefinite programming with a difference of convex function approach Rosa Which means they use an approximate method called Algorithm two to solve it because the problem gets complicated by its nonlinear cost function Taro <ref:2506.19829#pg2>

Paper summary: Rosa: The paper shows that these optimization problems can be framed in a way that's solvable, even if the original setup looks really messy with those bilinear constraints Dev It’s about finding a path to a solution when the math gets too complex for a direct solve Taro

Dev: And the simulation results on an aircraft are pretty telling. They show controllers that actively deteriorate adversarial observability while still maintaining near-optimal performance levels Rosa That means you get both goals somewhat satisfied at the same time Taro

Rosa: Plus, the analysis shows that as you make the control gain more flexible to deviate from perfect optimality, that control gain K hats increasingly decouples the first state of the system from the second one Dev Which is a kind of side effect we need to watch out for in deployment Taro

Taro: Decoupling states sounds like a potential failure mode if you're not careful about how much flexibility you allow in that gain setting Rosa And they also show that Algorithm one magnifies adversarial estimation errors across all states with almost no loss of performance Dev So, even when we get the observability reduction we want, the estimation error doesn't skyrocket Taro <ref:2506.19829#pg1>

Rosa: It sounds like a pretty robust result if those simulation results hold up under real-world conditions Dev But what's the big picture implication for someone who just wants to understand this concept?

Taro: It means there are fundamental limits to how much you can reduce observability given a set level of performance degradation Rosa And the paper validates that idea, showing that when minimizing the trace of the observability Gramian, smaller performance trade-offs give you a better payoff in reducing adversarial observability Dev

Dev: The lower bound for Problem two where they optimize the Gramian inverse, can actually become unbounded as a certain parameter lambda increases Taro That’s because a larger lambda allows that inverse to become singular Rosa It shows that the limits are not fixed, but depend on those specific parameters like kBk which dictates how much actuation authority you have for the operator Dev <ref:2506.19829#pg2>

Rosa: So, in short, they give us exact algorithmic solutions for the trace of the Gramian optimization and an approximate solution for the inverse using sequential semidefinite programming Taro It’s a practical way to handle this kind of control design problem Dev

Taro: The future work they mention is looking at strict output feedback controllers that only rely on partial state information but still manage to balance performance and observability reduction Rosa That suggests the goal isn't just about having perfect information, but managing what information you *do* have effectively Dev

Rosa: Right. So, this paper explores how we can deliberately design a system to be less observable by an adversary without completely breaking our control requirements Taro It’s about finding that precise balance between performance and security metrics Dev

Conclusion: Rosa: So, we're wrapping up this look at "Adversarial Observability and Performance Trade-offs in Optimal Control." Basically, they’re showing how you can actively make a system harder to observe by an adversary without messing up the actual driving performance.

Dev: Yeah, it boils down to finding that tight balance between keeping the control law effective and making sure the sensor measurements don't tell the enemy too much.

Rosa: They used these two metrics—the trace of the observability Gramian and its inverse—to show how much observability you can trade for performance degradation.

Taro: That’s interesting because they aren't just looking at one thing, they’re measuring both the average observability and those weird directions where the system is most hidden.

Dev: Right, and what really hits me is that they set some hard limits on this tradeoff based on how far you are from perfect control performance.

Rosa: They show that if you want to reduce observability, but keep your performance loss small, you get a better return in terms of how much harder it is for the adversary to see the system.

Taro: But they also flag some tricky things about this optimization process—specifically how the Gramian inverse can become singular if you push that tradeoff parameter too high.

Dev: That means there’s a point where the math breaks down, and you have to be careful with your control authority settings, like that kBk value they mentioned.

Rosa: So, the main thing is they give us these concrete ways to solve the optimization problem—one exact method for the average observability trace and an approximation for the inverse.

Taro: It gives us a blueprint for designing systems where security isn't just an afterthought, but something you actively engineer into the control structure itself.

University of Texas at Austin · Oden Institute for Computational Engineering & Sciences

eess.SY, cs.SY, math.OC

Submitted: 2025-06-24

Updated: 2026-11-10

Comments: 8 pages, 4 Figures

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints<ref:2506.19829#pg2>.

Key concepts

Observability Gramian Trace
This metric measures how well the system's states can be seen by sensors on average. A lower trace indicates poorer observability, meaning the system is harder to observe from an adversarial perspective.
Performance Trade-off Distance
This quantifies how far the controller's actual performance is from its ideal, optimal state. The paper analyzes how this distance affects the degree to which adversarial observability can be reduced.
Semidefinite Programming (SDP)
SDP is a mathematical tool used to solve complex optimization problems where both the objective function and constraints are convex. It allows for efficient computation when optimizing control laws or Gramian traces.
Observability Gramian Inverse Trace
This metric focuses on the directions in state space that are least observable. A higher trace here indicates a greater vulnerability to adversarial attacks along those specific, poorly observed state directions.

Terminology

Summary

We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints<ref:2506. p. 0

Where the paper says it

1 Adversarial Observability and Performance Trade-offs in Optimal Control Filippos Fotiadis1 , Member, IEEE, Ufuk Topcu1 , Fellow, IEEE Abstract—We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints. We quantify the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system.

19829#pg2>. The paper quantifies the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system<ref:2506. p. 0
Where the paper says it

1 Adversarial Observability and Performance Trade-offs in Optimal Control Filippos Fotiadis1 , Member, IEEE, Ufuk Topcu1 , Fellow, IEEE Abstract—We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints. We quantify the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system.

19829#pg2>. It derives theoretical lower bounds on these metrics under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off<ref:2506. p. 0
Where the paper says it

We quantify the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system. We derive theoretical lower bounds on these metrics under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off.

19829#pg2>. Finally, it shows that the performance-constrained optimization of the Gramian’s trace can be formulated as a one-shot semidefinite program, while it addresses the optimization of its inverse through sequential semidefinite programming<ref:2506. p. 0
Where the paper says it

We derive theoretical lower bounds on these metrics under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off. Finally, we show that the performance-constrained optimization of the Gramian’s trace can be formulated as a one-shot semidefinite program, while we address the optimization of its inverse through sequential semidefinite programming.

19829#pg2>.

Problem Setup and Metrics

The problem considers a system described by x(t) = Ax(t) + Bu(t)<ref:2506.19829#pg3>. The operator designs a control law Kx(t), Kp Rˆn, which stabilizes the system to the origin optimally and makes x(t) difficult to observe from y(t)<ref:2506. p. 1

Where the paper says it

We also assume that the operator of system (1), who designs the control input uptq, has access to the full state xptq. In light of this sensing asymmetry, the operator is motivated to design a control law uptq “ Kxptq, K P R mˆn , which does not only optimally stabilize system (1) to the origin, but also makes xptq difficult to observe from the output yptq.

19829#pg3>. To quantify adversarial observability, the authors adopt two metrics: i) the trace of the observability Gramian, which quantifies adversarial observability on average, across all observable directions<ref:2506. p. 0
Where the paper says it

We specifically quantify the observability of adversarial sensors by adopting the trace of the observability Gramian and its inverse. The former quantifies adversarial observability on average, across all observable directions, while the latter is skewed towards the less observable directions.

19829#pg2>, and ii) the trace of its inverse, which is skewed towards the less observable directions<ref:2506. p. 0
Where the paper says it

We specifically quantify the observability of adversarial sensors by adopting the trace of the observability Gramian and its inverse. The former quantifies adversarial observability on average, across all observable directions, while the latter is skewed towards the less observable directions.

19829#pg2>.

Trade-off Quantification

Theoretical lower bounds are established for these metrics as a function of the trade-off distance from optimal performance<ref:2506.19829#pg4>. The paper shows that smaller performance trade-offs generally yield a sharper decrease in the trace of the observability Gramian, whereas larger performance trade-offs lead to a steeper increase in the trace of the Gramian’s inverse<ref:2506. p. 0

Where the paper says it

For both of these metrics, we establish theoretical lower bounds on their optimal values as a function of the trade-off distance from optimal performance. We show that smaller performance trade-offs generally yield a sharper decrease in the trace of the observability Gramian, whereas larger performance trade-offs lead to a steeper increase in the trace of the Gramian’s inverse.

19829#pg4>. Specifically, smaller performance losses yield steeper trade-offs in reducing adversarial observability for the cost (12)<ref:2506. p. 3
Where the paper says it

λq for small λ. Looking at (12), this implies that small performance losses λ yield steeper trade-offs in reducing adversarial observability for the cost (12).

19829#pg5>.

Optimization Approaches

The optimization of the trace of the Gramian is shown to boil down to solving a semidefinite program (SDP)<ref:2506.19829#pg5>. Conversely, the optimization of the trace of the Gramian inverse is addressed approximately through a sequential SDP algorithm that is obtained by applying the convex-concave procedure<ref:2506. p. 0

Where the paper says it

We show that smaller performance trade-offs generally yield a sharper decrease in the trace of the observability Gramian, whereas larger performance trade-offs lead to a steeper increase in the trace of the Gramian’s inverse. Finally, we solve the constrained optimization of these metrics i) precisely for the trace of the Gramian, by proving it boils down to the solution of a semidefinite program (SDP); and ii) approximately for the trace of the Gramian inverse, through a sequential SDP algorithm that we obtain by applying the convex-concave procedure.

19829#pg5>. Problem 1 is reformulated as an SDP where both the objective and constraints are convex, enabling efficient computation<ref:2506. p. 4
Where the paper says it

Theorem 2 enables us to reformulate the original Problem 1 as an SDP where both the objective and constraints are convex. This convex formulation preserves the structural properties of the original problem while allowing for efficient computation using off-the-shelf SDP solvers.

19829#pg6>. Problem 2, involving the trace of the Gramian inverse, is solved approximately using Algorithm 2, which relies on a difference of convex functions (DC) program with a linear cost function<ref:2506. p. 4
Where the paper says it

Specifically, we show Problem 2 can be cast as an equivalent difference of convex functions (DC) program with a linear objective function that involves no inverses of matrix variables. Under the transformation Yǫ “ W´1 ǫ , Problem 2 admits the same optimal solution as the DC program with linear cost: min K,P,Yǫ ´ trpYǫV ´1 q (24) s.

19829#pg7>.

Performance and Robustness Analysis

Simulations on an aircraft demonstrate that the proposed scheme yields controllers that deteriorate adversarial observability while having near-optimal performance<ref:2506.19829#pg8>. The analysis shows that as the gain becomes more flexible to deviate from strict optimality, the resulting control gain Kˆ increasingly decouples the first state of the system from the second one<ref:2506. p. 6

Where the paper says it

0018 , whereas the nominal optimal closed-loop matrix A ` BK‹ is: A ` BK‹ “ „ ´0. The pattern we notice is the one we would ideally expect: as we gain more flexibility to deviate from strict optimality, the resulting control gain Kˆ increasingly decouples the first state of the system from the second one.

19829#pg8>. Furthermore, Algorithm 1 is shown to magnify adversarial estimation errors across all states with almost negligible loss of performance. p. 7
Where the paper says it

4: Trajectories of the adversary’s observation errors, and the closed-loop performance cost. states with almost negligible loss of performance, while Algorithm 1 increases the estimation error of the fourth state to completely unreliable levels with a little more performance trade-off.

Conclusion

The paper concludes by establishing lower bounds for Problems 1 and 2 across all values of the trade-off parameter λ that quantify the fundamental limits of adversarial observability reduction under a given level of performance degradation. The results validate the intuition behind Theorem 3, showing that when minimizing the trace of the observability Gramian, smaller performance trade-offs yield a larger payoff in adversarial observability. p. 6

Where the paper says it

While they are relatively tight for Problem 1, they become looser for Problem 2 because the inverse of the Gramian tends to be close to singular as λ increases, and because the obtained solution is only a local minimum. Finally, we also validate the intuition behind Theorem 3: when minimizing the trace of the observability Gramian, smaller performance trade-offs yield a larger payoff in adversarial observability.

The lower bound in Problem 2 can become unbounded as λ increases, since a larger λ allows the Gramian inverse to become singular. p. 3
Where the paper says it

Looking at (12), this implies that small performance losses λ yield steeper trade-offs in reducing adversarial observability for the cost (12). On the other hand, the lower bound in Problem 2 can become unbounded as λ increases, since a larger λ allows the Gramian inverse to become singular.

The lower bounds also depend on other parameters, most notably kBk, which dictates that larger actuation authority for the operator allows for sharper decrease in adversarial observability. p. 3
Where the paper says it

The lower bounds also depend on a number of other parameters. Most notable is kBk, which dictates that larger actuation authority for the operator allows for sharper decrease in adversarial observability.

The paper provides an exact algorithmic solution for Problem 1 and an approximate algorithmic solution for Problem 2 through sequential SDP<ref:2506. p. 4
Where the paper says it

ALGORITHMIC SOLUTIONS TO PROBLEMS 1 AND 2 We subsequently provide algorithmic solutions to Problems 1 and 2. Specifically, we provide an exact algorithmic solution for Problem 1 by reformulating it into an equivalent SDP, and an approximate algorithmic solution for Problem 2 using sequential SDP.

19829#pg5>. The future work includes the consideration of strict output feedback controllers that rely only on partial state information, but which still achieve a balance between having sufficient performance and minimizing adversarial observability. p. 7
Where the paper says it

Finally, we optimize the trace of the observability Gramian through an SDP, while, for the trace of the Gramian inverse, we perform this optimization approximately using sequential SDP. Future work includes the consideration of strict output feedback controllers that rely only on partial state information, but which still achieve a balance between having sufficient performance and minimizing adversarial observability.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adversary Observability Minimization via Optimal Control: The system can be designed to actively reduce an adversary's ability to observe a system by designing a feedback gain K such that pA BK C is as unobservable as possible while adhering to strict closed-loop performance constraints.

  2. Quantifying Trade-offs with Gramian Metrics: The AI system can quantify adversarial observability using the trace of the observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system.

  3. Theoretical Lower Bounds for Performance-Constrained Observability: The system can establish theoretical lower bounds on metrics like the trace of the Gramian under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off.

  4. Exact Optimization via Semidefinite Programming (SDP): The controller design problem that minimizes the trace of the observability Gramian can be formulated and solved precisely as a one-shot semidefinite program.

  5. Approximate Optimization for Least Observable Directions: The optimization of the trace of the Gramian inverse, which is more complex, can be addressed approximately through a sequential semidefinite programming algorithm using a convex-concave procedure.

  6. Robustness to Sensing Uncertainty: The AI system can provide certified upper bounds on adversarial observability metrics even when the operator only knows a superset of the adversary's possible sensing configurations.

Abstract

We develop a feedback controller that minimizes the observability of a set of adversarial sensors of a linear system, while adhering to strict closed-loop performance constraints. We quantify the effectiveness of adversarial sensors using the trace of their observability Gramian and its inverse, capturing both average observability and the least observable state directions of the system. We derive theoretical lower bounds on these metrics under performance constraints, characterizing the fundamental limits of observability reduction as a function of the performance trade-off. Finally, we show that the performance-constrained optimization of the Gramian's trace can be formulated as a one-shot semidefinite program, while we address the optimization of its inverse through sequential semidefinite programming. Aircraft and satellite simulations demonstrate that the proposed controllers reduce adversarial observability while maintaining near-optimal performance.

Related papers