Skip to content

Latest commit

 

History

History
150 lines (103 loc) · 12.5 KB

File metadata and controls

150 lines (103 loc) · 12.5 KB

Personal Research Directions

Research identity

Build decision-centred cyber and network systems research in which an operational security or social-system problem is converted into a closed, auditable chain:

mechanism → intervention/strategy → formal result → verified computation → bounded decision insight

Prefer a narrow, defensible contribution with complete theory–code–evidence alignment over a broad framework assembled from fashionable components.

Portfolio maturity

Maturity Directions Interpretation
Established publication core propagation dynamics, threshold/stability analysis, malware/APT defense, optimal control, differential and strategic games Use as the default intellectual foundation and closest self-prior-work boundary.
Current expansion impulse/hybrid control, cyber propaganda and rumor, awareness/opinion coupling, social-engineering malware, employee security compliance, hierarchical incentive/audit design, cost-effective response Treat as active combinations with meaningful recent corpus support.
Emerging or repository-derived capability topology-first node simulation, RL/MARL/MPC, attack-graph hybrids, PINN/PIDL, missing-mechanism learning Add only when the question and data regime require them; do not imply equal corpus maturity.
Evidence-enabling layer observed topology, traces, calibration, emulation, cyber range, hardware-in-loop Use to close a named validity gap rather than as a standalone novelty claim.

Core direction portfolio

1. Cyber propagation, containment, and resilience

Study malware, mobile/IoT threats, APTs, social engineering, cyberbullying, rumor, propaganda, awareness, patching, quarantine, backup, and recovery as dynamical decision problems.

  • Central question: how do propagation mechanisms, topology, human response, and resource limits change the timing and allocation of defense?
  • Preferred formal objects: compartmental or node-level network dynamics, epidemic/awareness coupling, attack graphs, continuous–impulsive systems.
  • Minimum evidence: invariants or well-posedness, verified solver, no-defense and simple-policy baselines, topology/parameter sensitivity, bounded security interpretation.
  • Boundary: stylized dynamics do not establish observed attack rates or deployment effectiveness.

2. Optimal, impulse, and hybrid cyber control

Design cost-effective continuous, sampled, impulse, and hybrid interventions under explicit budgets, timing, and admissibility constraints.

  • Central question: when and how should limited security resources be allocated?
  • Preferred formal objects: optimal control, Pontryagin systems, impulse timing, hybrid reset maps, direct transcription or forward–backward sweep.
  • Minimum evidence: necessary-versus-sufficient status, stationarity/residual checks, independent numerical comparison, grid/tolerance sensitivity, matched budgets.
  • Boundary: a converged numerical trajectory is a configured solution, not a global optimality certificate.

3. Differential, Stackelberg, coalition, and repeated games

Model adaptive attackers, defenders, service providers, organizations, employees, or social actors with explicit information and strategy classes. The current corpus supports a research-architecture branch that couples organizational incentives and audits, employee compliance responses, and node-level intranet risk in a Stackelberg differential-game formulation.

  • Central question: how do strategic timing, leadership, cooperation, or deviation incentives reshape cyber defense?
  • Preferred formal objects: open-loop or feedback differential games, Stackelberg games, bilevel leader–follower optimization, FlipIt-style timing, coalition or partition stability, repeated/self-play interaction.
  • Minimum evidence: declared solution concept, admissible strategies, response matrix, unilateral deviations or independently solved best responses, exploitability/stability diagnostics.
  • Boundary: tested local deviations do not establish a general Nash, Stackelberg, or coalition-stability theorem.

4. Topology, heterogeneity, and network-scale intervention

Make node/edge semantics and structural heterogeneity part of the scientific question rather than an implementation detail.

  • Central question: which conclusions depend on hubs, communities, direction, weights, temporal structure, or scale?
  • Preferred formal objects: node-level and degree-class models, heterogeneous rates/costs, multilayer or temporal graphs, topology-aware controls.
  • Minimum evidence: a topology card, deterministic preprocessing, multiple graph instances, matched structural controls, perturbation/rewiring tests, held-out graphs for generalization.
  • Boundary: observed topology plus assumed dynamics is real-topology simulation, not real-world validation.

5. Feedback learning and multi-agent cyber control

Add RL, MARL, CTDE, self-play, or model-predictive feedback only when sampled observations, adaptation, or unknown opponents are central.

  • Central question: can a feedback policy improve decisions under partial observation, stochasticity, or strategic adaptation?
  • Preferred formal objects: MDPs, Markov games, hybrid environments, actor–critic or value-based policies, self-play and response analysis.
  • Minimum evidence: parity with the canonical continuous model, exact action/event/reward timing, physical metrics beyond reward, matched information/budget/tuning, repeated held-out scenarios, failure retention.
  • Boundary: reward improvement does not establish physical effectiveness, equilibrium, robustness, or field performance.

6. Physics-informed inverse problems and missing-mechanism learning

Use PINN, PIDL, inverse estimation, or PMP-informed learning when states are sparse/noisy, parameters are unknown, or a mechanistic component is missing.

  • Central question: what exact state, parameter, function, or correction is unknown, and is it identifiable from the observation design?
  • Preferred formal objects: inverse PINNs, physics-informed dynamic learning, learned corrections, neural control constrained by optimality systems.
  • Minimum evidence: explicit observation model, separate data and collocation points, generated-truth recovery tests, held-out nodes/times/graphs, misspecification baselines, independent rollout, identifiability analysis.
  • Boundary: residual consistency is not parameter truth; real-data fitting is calibration or effective-rate estimation unless ground truth and identifiability are established.

7. Evidence expansion from models to operational systems

Strengthen mechanistic papers through lawful observed topology, traces, calibration, emulation, cyber ranges, or hardware-in-the-loop only when they close a named validity gap.

  • Central question: what stronger claim becomes supportable after adding this evidence source?
  • Preferred progression: Track A verified numerics → Track B closed-loop simulation → Track C topology/trace evidence → Track D calibration/prediction → Track E measured feasibility.
  • Minimum evidence: provenance, licences, data-to-model mapping, splits/leakage checks, uncertainty, failure cases, and a reality ledger.
  • Boundary: adding more evidence types is not automatically stronger; select the smallest adequate track.

Layer-selection doctrine

Start from the Foundation layer. Add another layer only when its scientific trigger is present.

Trigger Add Do not add merely for
Canonical dynamics, control/game formulation, verified numerics Foundation every project already starts here
Sampled feedback, stochastic operation, adaptive opponent, repeated interaction Game learning fashion, algorithm count, or a stronger-looking title
Partial/noisy observations, unknown parameters, missing mechanism, learned optimality system Physics informed replacing a solvable mechanistic model with a neural surrogate
Observed topology or trace closes a structural/temporal validity gap Track C calling synthetic dynamics real-world data
Real observations support calibration and held-out prediction Track D claiming true parameters without identification
Testbed or field measurement supports feasibility Track E implying broad deployment or causal effectiveness

When combining layers, preserve one state order, units system, adjacency convention, action timing, reset map, objective, and reality classification.

Direction selection gates

Score a proposed direction only after a closest-full-text audit.

  1. Decision value: identify the actor, action, cost, uncertainty, and consequence.
  2. Exact gap: state what the closest work cannot represent, solve, verify, or observe.
  3. Minimal mechanism: isolate the smallest interaction that creates the contribution.
  4. Formal tractability: define a plausible theorem, optimality, equilibrium, or estimation ladder.
  5. Evidence feasibility: identify the smallest adequate track and a decisive experiment.
  6. Asset fit: reuse the three repository layers through pinned interfaces without treating them as novelty.
  7. Journal fit: make the paper's center of gravity match TIFS, TDSC, TSMC, TNSE, or TCSS.
  8. Kill criteria: name prior-art, proof, code, data, identifiability, or effect-size outcomes that would stop or narrow the direction.

Prefer a direction when all eight gates have defensible answers. Otherwise keep it as exploratory or record a NO-GO decision.

Journal alignment

  • TIFS: lead with adversary capability, security consequence, defender decision, and security-specific evaluation.
  • TDSC: lead with dependable/secure system behavior, fault/threat scope, guarantees, overhead, and failure modes.
  • TSMC: lead with integrated system dynamics, stability/feasibility, human–machine or cyber-physical interaction, and system interpretation.
  • TNSE: lead with network representation, topology, heterogeneity, temporal structure, scale, and network-aware intervention.
  • TCSS: lead with computational social mechanisms, awareness/opinion/incentives, ethical limits, and policy interpretation.

Verify current official submission rules before final formatting; these mappings describe research fit, not cached author instructions.

Preferred contribution architecture

Use two to four dependent contributions:

  1. formulate the decision problem and model;
  2. characterize the formal solution or conditions;
  3. develop the reproducible solver, estimator, or policy;
  4. provide the evidence design and bounded finding.

Do not split one idea into decorative contributions. Do not claim a standard dataset, public topology, tutorial method, or generic neural component as novelty without a defensible new mapping or evaluation design.

Personal execution preferences

  • Treat supplied research bundles as working packages: unpack, locate the real workspace, and read its current routing files before acting.
  • Reuse a supplied framework when it is a good foundation; do not rebuild from scratch without a reason.
  • Default to scope lock → audit → validation → execution for a new “direction.”
  • Perform writing, derivation, experiments, simulation, and release checks as an end-to-end chain when the request calls for them.
  • Preserve adverse literature, failed runs, narrowed claims, pending gates, and NO-GO outcomes.
  • Keep authenticated full-text provenance and access boundaries explicit; abstract-only review is not equivalent to full-text verification.
  • Validate shipped archives and clean re-extractions, not only the working tree, before claiming a release is ready.

Current provenance boundary

The profile has expanded-partial-calibration: 73 verified full texts were deep-read for research architecture and correction doctrine, comprising 42 owner-provided papers and 31 nonduplicate open-access journal articles. Sentence-level style remains calibrated from the role-weighted local corpus, including 20 first/corresponding-author core-voice records and 17 IEEE Transactions papers. Corresponding status is a useful authorship signal but does not prove section-level drafting ownership; many contribution roles remain unconfirmed. The package records lawful metadata, aggregate counts, derived metrics, rules, and paper cards but no article full text or per-paper deep-reading notes. All 17 owner-priority selected Transactions rows have private full text; one different-subfield row is explicitly owner-deprioritized. The broader publication universe, section-level voice attribution, venue-specific microstyle, and publisher-final verification of the 2026 TDSC-matched manuscript remain open; do not use full-corpus-calibrated until the full audit protocol passes.