When a crystal is stretched, some of its vibration patterns weaken. Physicists find which ones by computing the lattice's vibration modes, the collective patterns in which atoms oscillate together, and tracking the stiffest restoring force against each. The pattern whose restoring force falls toward zero is the "soft mode." It shows where the lattice will give way, and a standard derivative identity attributes that softening to individual bonds. This note proposes using the same machinery to decide how an automated network defender should cut, redirect, and quarantine traffic. The defender writes the attacker's spread through the network as a linear system, finds the softest mode of that system, and takes the cheapest set of network actions that keeps this mode's restoring force above a chosen safety margin. It does not stop at merely positive, and it does not push the margin as high as possible.
The opportunity is practical. Learned defenders score well on the attackers they were trained against and collapse on others. In Autonomous Network Defence using Reinforcement Learning, a specialist agent scores -61.05 against an unfamiliar attacker, worse than random at -37.6. These defenders are also easy to poison: in Reinforcement Learning for Autonomous Defence in Software-Defined Networking, preserved hosts fall from about 20 to 3.3 under state manipulation. A containment rule derived from the network's structure needs no training against any particular attacker. It gives an interpretable certificate for each action and has one knob that trades security against availability. The question is whether it is enough to match learned agents in-distribution and beat them out of distribution.
Core Claim
Let $w$ be the vector of per-link and per-host weights the defender controls through software-defined networking (SDN), where a central controller rewrites switch forwarding rules. Each weight is 1 when traffic is allowed, 0 under Drop or Isolate, and a reduced value under rate limiting or Reflect-to-honeypot. Define the containment stiffness matrix
$$K(w) = \Delta - B(w),$$
where $\Delta$ is a diagonal matrix of per-host clean-up rates (restore or remove actions) and $B(w)$ is the weighted matrix of attacker reachability. Its entry $B_{ij}$ equals $w_{ij}$ times the estimated probability that a foothold on host $j$ yields one on host $i$ in a time step.
Linearizing a susceptible-infected-susceptible spread model around the "no compromise" state shows that footholds die out when the smallest eigenvalue $\mu_{\min}(K)$ is positive and grow along the eigenvector $v$ when it is negative. This is the network analogue of a phonon mode going from real to imaginary frequency. Following the Hellmann-Feynman logic that attributes phonon softening to individual force constants, the sensitivity of the soft mode to each action is
$$\frac{\partial \mu_{\min}}{\partial w_e} = v^{\top}\,\frac{\partial K}{\partial w_e}\,v,$$
so each candidate action is scored by how much it stiffens the soft mode. The defender solves
$$\min_{a}\; C_{\text{avail}}(a) \quad \text{s.t.} \quad \mu_{\min}\big(K(w(a))\big) \ge \varepsilon,$$
where $a$ is a set of SDN actions, $C_{\text{avail}}$ is the availability cost of those actions (blocked legitimate flows, hosts taken down), and $\varepsilon \ge 0$ is the stability margin. The optimization is solved greedily using the gradient above, then re-solved each step as alerts update $B$.
The claim is that this structural controller, run with no attacker-specific training, matches hierarchical reinforcement learning (RL) on in-distribution attackers in CybORG. It should substantially beat any single learned policy on mismatched attackers. It should also degrade gracefully, not catastrophically, under the state-poisoning attacks of Han et al.
Where the Physics Actually Enters, and Where It Does Not
The strain literature supplies method and one design lesson. It does not supply a shared phenomenon, and the intersection file confirms no prior work links the two areas. The method is mode analysis. Kong et al. (Superconductivity in electron-doped arsenene) and Qiu et al. (Superconductivity in monolayer Ba2N electride; Superconductivity in monolayer Ba 2 N electride: First-principles study) each localize a strain response to a single softened mode at the K point and to the bonds that drive it. That workflow, finding the soft eigenvector and attributing it to local stiffnesses, is what the controller reuses.
The design lesson comes from Hsu et al. (Manipulating superconductivity in ruthenates through Fermi surface engineering). They find the best operating point lies near, not on, a singularity, because sitting on it triggers competing instabilities. The containment analogue runs in two directions. At $\varepsilon = 0$ the defender sits on the threshold, where model error flips the sign of $\mu_{\min}$. At very large $\varepsilon$ the defender over-isolates and causes its own denial of service, the failure mode the CAGE penalty structure and FedDICE's clinical setting (FedDICE) both punish. The margin $\varepsilon$ is the knob that sweeps between them.
Hofmann et al. (Heuristic bounds on superconductivity and how to exceed them) prove that coupling concentrated in one mode drives collapse, while coupling spread over many modes does not. This gives a testable prediction rather than a design rule. Attackers that concentrate on one bridging path, like the shortest-path B_line attacker, should produce a sharply localized soft eigenvector that a cheap cut can stiffen. Exploratory attackers like RedMeander should produce delocalized soft modes that need more distributed, costlier action. The inverse participation ratio of $v$, a standard measure of mode localization, should predict containment cost.
The proposal is honest about what is old. Spectral thresholds for epidemic spread on graphs are well known in network science, though the file contains no instance of them in automated containment. What is new here is margin-controlled, availability-costed containment executed through concrete SDN primitives, evaluated against learned defenders on generalization and poisoning.
Research Program
The first step is to build the stiffness matrix from what a defender actually observes. The topology comes from the environment. Reachability probabilities come from an attacker-agnostic prior, namely uniform exploit probability over exposed services, so the controller never sees attacker-specific data. The diagonal clean-up rates come from the defender's restore and remove actions. Current alerts raise the belief that particular hosts are compromised, which acts as a source term and reweights $B$.
The second step maps actions onto weight changes. Drop and Isolate set weights to zero. Reflect sends the attacker's flow to a host with zero onward reachability and zero availability cost to legitimate users. These are the primitives of SecControl (Bridging the Gap Between Security Tools and SDN Controllers), which measured switch deployment latencies of 11.2 ms for Drop and 17.6 ms for Reflect. Those latencies bound how fast the controller's decisions can take effect in an emulated network.
The third step studies the controller's two failure surfaces. One is sensitivity to a misspecified $B$, tested by perturbing the reachability prior. The other is sensitivity to poisoned observations, tested by flipping link and host status bits as in Han et al. Corrupted inputs can still mislead a structural controller. The question is whether the damage is bounded, because one flipped bit perturbs $\mu_{\min}$ by at most the corresponding entry's weight under $v$.
Baselines
The naive reference points are the random "chance" defender in CAGE Challenge 1, which Foley et al. report at an average score of -37.6, and a do-nothing defender. The proposal must clear both by a wide margin on every attacker.
The strongest existing method is the hierarchical PPO agent of Foley et al., which won CAGE Challenge 1 with an average score of -8.43 in the CybORG environment. It is the right competitor because it already handles two attacker types, through a learned controller that identifies attacker intent. It will be retrained from released settings on the same 13-host topology with the same episode lengths and reward. It will be compared with the soft-mode controller, which gets an equal compute budget for its single tuned hyperparameter $\varepsilon$, selected on held-out seeds of the training attacker only.
Beating this baseline in-distribution means a score within one standard error of -8.43 or better. The real test is a held-out attacker condition. There the RL agent is trained on one red agent only and evaluated on the other, which is the regime where Foley et al. report -61.05 for a specialist. Beating it there means a clearly better score than the mismatched RL agent at matched seeds.
On the 32-node Mininet/OpenDaylight testbed of Han et al., the comparison points are DDQN and A3C. They preserved about 20 to 21 hosts clean and 3.3 (DDQN) and 4.9 (A3C) under state manipulation. Beating them means matching the clean count and preserving substantially more hosts under the identical poisoning schedule.
The proposal also needs a rule-based reference: Snort-style thresholds triggering Drop. It needs a coarse reference: quarantine the entire subnet of any alerting host. For ARCS (ARCS), only its own dataset metrics are reported, such as 89.2% response effectiveness and 23.5 s response time. No reported result was found on Lacuna for ARCS in CybORG, so it serves as context, not a head-to-head number.
Three ablations isolate the key ingredient. The first replaces eigenvector attribution with degree-based or betweenness-based action scoring under the same margin constraint, testing whether the soft mode matters or any centrality would do. The second fixes $\varepsilon = 0$, testing the "near but not on" lesson. The third drops the availability cost, testing whether the constrained form or plain spectral minimization does the work.
Why not just…?
Why not just use the hierarchical RL agent, which already wins CAGE? Its controller generalizes only across the attackers it was trained on. Foley et al. show that a component trained on one attacker scores -61.05 against another. No poisoning evaluation of it appears in the file, while Han et al. show learned defenders of this kind lose most of their protected hosts under state flipping. If the hierarchical agent matches the soft-mode controller on held-out attackers and under poisoning, the proposal's value reduces to interpretability. The held-out-attacker and poisoning experiments are designed to settle exactly this.
Why not just isolate the highest-centrality hosts or cut the shortest path to the critical server? This is the most serious threat. On a 13-host, three-subnet network, the soft eigenvector may simply coincide with the shortest path, making the spectral machinery unnecessary. The centrality ablation tests this directly. The scaled experiments on the 32-node testbed and on larger generated topologies with redundant paths test the regime where the soft mode should diverge from any single path. If centrality scoring matches eigenvector scoring there as well, the honest conclusion is that the margin constraint, not the mode analysis, carries the result.
Why not just quarantine the whole affected subnet? It is fast and needs no model, but CAGE penalizes disrupted services. FedDICE shows that in clinical networks, wholesale isolation disconnects life-critical devices, which is why it uses device-level rules. The full-quarantine baseline will be plotted on the same compromise-versus-availability frontier traced by sweeping $\varepsilon$. The proposal wins only if its frontier dominates.
Why not just adversarially train the RL agent against poisoning? This is plausible and untested in the file. It still ties robustness to the poisoning distribution seen in training. The comparison will include an RL agent trained with Han et al.'s flipping attack, evaluated under a different flipping rate than it saw.
Proposed Experiments
The decisive check has three parts, each with a prediction that can fail.
First, the controller's diagnostic must carry information. At each step, $\mu_{\min}$ computed from the defender's belief should predict compromise of the operational server within the next few steps better than alert count or a centrality-weighted alert sum. The comparison uses the area under the ROC curve, with the window matching the two-to-four-alert window reported by Cao et al. (Security Testbed for Preempting Attacks against Supercomputing Infrastructure). If $\mu_{\min}$ is no better than alert count, the mode picture adds nothing.
Second, the generalization gap must reverse. In CybORG, train on B_line, test on RedMeander, and the reverse. The soft-mode controller needs no attacker training, so its in-distribution and out-of-distribution scores should be close, while the RL specialist's should diverge as Foley et al. report. The localization prediction from Hofmann's argument is also checked here: RedMeander episodes should show more delocalized soft modes and higher containment cost.
Third, under Han et al.'s state-manipulation schedule on the 32-node testbed, the soft-mode controller's preserved-host count should degrade roughly in proportion to the number of flipped bits rather than collapsing.
Sweeping $\varepsilon$ yields the compromise-versus-availability frontier. Actions are replayed through SecControl-style primitives to confirm that per-step decisions fit within the measured millisecond deployment latencies.
References
- Autonomous Network Defence using Reinforcement Learning
- Reinforcement Learning for Autonomous Defence in Software-Defined Networking
- Superconductivity in electron-doped arsenene
- Superconductivity in monolayer <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msub><mml:mi>Ba</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:mi mathvariant="normal">N</mml:mi></mml:mrow></mml:math> electride: First-principles study
- Manipulating superconductivity in ruthenates through Fermi surface engineering
- FedDICE: A Ransomware Spread Detection in a Distributed Integrated Clinical Environment Using Federated Learning and SDN Based Mitigation
- Heuristic bounds on superconductivity and how to exceed them
- Bridging the Gap Between Security Tools and SDN Controllers
- CybORG: An Autonomous Cyber Operations Research Gym
- ARCS: Adaptive Reinforcement Learning Framework for Automated Cybersecurity Incident Response Strategy Optimization
- Security Testbed for Preempting Attacks against Supercomputing Infrastructure