When to Deceive: A Cross-Layer Stackelberg Game Framework for Strategic Timing of Cyber Deception

Quanyan Zhu; Ya-Ting Yang

REVIEW 3 major objections 3 minor 3 cited by

When to Deceive: A Cross-Layer Stackelberg Game Framework for Strategic Timing of Cyber Deception

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Strategically timing cyber deception within a two-layer Stackelberg game reduces asset compromise risk by 25–50% relative to fixed schedules.

desk verdict A genuinely new two-layer deception-timing framework whose headline improvement is computed against an attacker who cannot act on his own belief update. read the letter →

arxiv 2505.21244 v1 pith:MZXRJIYU submitted 2025-05-27 cs.GT

Ya-Ting Yang , Quanyan Zhu This is my paper

classification cs.GT MSC 91A6591A1591A80

keywords cyberdeceptionStackelberggameMarkovstopping-timedecisionprocessdynamicprogrammingbeliefupdatemovingtargetdefenseadvancedpersistentthreats

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Cyber deception tools like honeypots and decoys lose their power once an attacker recognizes them, so a defender must decide not only which technique to use but when to switch. This paper claims that the timing and selection of deception should be treated as a two-level game: a tactical level where attacker and defender interact under one-sided information, and a strategic level where the defender chooses stopping times to switch deception modes within a budget. The paper shows in an enterprise-network simulation that such strategically timed switching improves the defender's expected utility and lowers the probability of critical-asset compromise by 25 to 50 percent relative to fixed, no-switch, or uniform-interval baselines. If true, this gives security teams a principled way to allocate limited deception resources dynamically instead of on a fixed schedule.

What carries the argument

The load-bearing mechanism is the separation of timescales into two coupled optimization problems. The tactical layer is a Markov game with one-sided information, in which the defender knows the deception mode $\theta$ and the attacker does not, and in which the attacker updates a belief $b_k(\theta)$ by Bayes' rule; its equilibrium is a one-sided $\epsilon$-Perfect Bayesian Nash Equilibrium. The strategic layer is a stopping-time decision process whose state is $\bar{x}_k=(s_k,b_k,\theta_k,m,\Theta')$, where the defender compares the value of staying in the current mode with the value of switching to any unused technique and switches only when the latter is larger, subject to a total budget of $M$ switches. Dynamic programming on this extended state, with belief updates computed from the tactical equilibrium, produces the optimal switching times and choices.

What would settle it

Rerun the Section IV enterprise-network experiments with an attacker strategy that conditions on the posterior belief (for example, choosing the action with highest expected reward under the belief) and compare the optimal switching policy against the no-switch baseline; if the margin shrinks below the reported 25–50% or reverses, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the question of when to deceive can be solved as a bi-level Stackelberg game, and that the resulting optimal switching policy materially outperforms static deception schedules. In the lower layer, for each deception mode the interaction is a one-sided information Markov game in which the defender knows the active mode but the attacker only observes the system state and maintains a Bayesian belief about it; the solution is a one-sided epsilon-Perfect Bayesian Nash Equilibrium computed by dynamic programming. In the upper layer, the defender's value function over an extended state consisting of the system state, the attacker's belief, the current mode, the remaining switches, and the used techniques dictates whether to stay with the current deception or switch to an unused one. Numerical experiments with a five-step attack path show that as the attacker's lifetime grows, the defender's advantage erodes without switching, while switching restores it; with a budget of one or two switches, the proposed strategy outperforms no-switch, fixed-timing, and uniform-interval baselines and reduces the risk of asset compromise by 25 to 50 percent relative to baselines without switching.

Load-bearing premise

The timing policy is evaluated against an attacker whose actions do not depend on the belief he is modeled as forming; a real attacker who used that belief to guide his actions could reduce or erase the reported 25–50% gain.

Editorial extensions

If this is right

Without switching, the defender's initial advantage erodes as the attacker's lifetime grows; strategic switching restores it by changing the game mode before the attacker's belief matures.
Each additional budgeted switch raises the defender's value at the initial stage for every tested attacker lifetime in the numerical study.
When one switch is allowed, the optimal policy switches to the sophisticated deception at stage 1 or 2, outperforming fixed-timing baselines, including one that switches to the wrong technique.
Uniform-interval switching performs worse than optimal switching because early or late activation lets the attacker settle his belief and press toward the critical asset.
Compared with no-switch baselines, strategically timed switching reduces the risk of asset compromise by 25–50% in the simulated enterprise network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension is a belief-threshold switching rule that triggers a switch when the attacker's posterior for the active mode crosses a threshold, which would make the policy implementable without solving the full dynamic program online.
Because the model treats each deception as a one-shot resource whose surprise value decays, the same timing logic could apply to other security controls with one-time use, such as credential rotation, decoy accounts, or re-keying, with different transition and reward parameters.
The 25–50% improvement is an upper bound under a passive-belief attacker; an attacker who best-responds to his posterior belief would likely shrink the margin, and testing that variant in the same simulation would reveal how much of the gain is timing versus belief restriction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, this one is worth a look but needs revision before the numbers can be trusted.

The new piece is the two-layer decomposition: a one-sided Markov game for a fixed deception, coupled to a stopping-time optimal switching problem for when to rotate deceptions. That combination is not in the cited literature, and the dynamic programming recursion in (8)-(9) is clean and internally consistent. The authors also correctly recognize that deception decays as the attacker learns, so timing matters. That is a real contribution to the cyber-deception subfield.

The load-bearing flaw is the attacker model. In Definition 1, the attacker updates a belief b_k by Bayes' rule, but his strategy is restricted to Markov strategies on the observed state only: pi_a: S -> Delta(A_a). Sequential rationality in Definition 3 optimizes over exactly that restricted class, so b_k never enters his action choice. The defender's switching policy, however, is explicitly built around keeping b_k below a threshold eta in (4). So the 25-50% improvement in the abstract is the value against an attacker who is barred from reacting to the strategic timing at all. That is not a full Stackelberg equilibrium; it is a constrained best response. The stress-test note is correct, and it lands.

Other issues are minor by comparison: the numerical section never states eta, the 25-50% figure is not tied to a specific row or table, Fig. 3 reports 200 experiments without error bars, and the parameters alpha, delta, beta are given only in the table caption. All fixable.

The paper is not a waste of time. The framework is thoughtfully put together, and the mathematical machinery is standard. The right fix is to let the attacker condition on his belief, even approximately, or to argue why state-only strategies are the correct information restriction. With that, the timing result would be meaningful. As it stands, the headline claim overreaches.

I'd send this to a serious referee. The model and the gap both deserve careful discussion. I probably wouldn't cite the numerical claim until the belief issue is resolved.

Referee Report

3 major / 3 minor

Summary. The paper proposes a cross-layer Stackelberg game framework for cyber deception timing. The lower tactical layer models defender-attacker interaction under a given deception as a one-sided information Markov game, where the defender knows the active deception mode but the attacker only observes the system state and maintains a belief over modes. The upper strategic layer is a stopping-time decision process that selects when and which deception technique to activate under a budget, using dynamic programming over an extended state that includes the current mode, remaining switches, and the attacker's belief. The authors claim that strategically timed deceptions improve defender utility and reduce asset compromise risk by 25-50% compared to fixed-schedule baselines. Numerical experiments on an enterprise network scenario illustrate the value of switching, with comparisons against no-switch, fixed-timing, and uniform-interval baselines.

Significance. If the modeling gaps were resolved, this framework would be a useful contribution to the cyber-deception timing literature: it explicitly couples a tactical Markov game with a strategic switching decision, and the numerical comparison against fixed-schedule baselines is a sensible evaluation design. The dynamic programming recursions in (8)-(9) are standard and internally consistent given the stated model, and the paper provides pseudocode (Algorithm 1) that conveys the intended execution. However, the central quantitative claim of a 25-50% improvement is not currently supported by any specific numerical result, and the model contains a load-bearing inconsistency between the attacker's Bayesian belief and his restricted strategy class, as well as between the fixed-mode belief update and the time-varying mode induced by switching. These issues prevent the paper from being accepted in its present form.

major comments (3)

[Section II-A, Definition 3, Eq. (1)] The attacker's strategy is restricted to Markov strategies of the form π_a^k: S → Δ(A_a), so his action cannot depend on the belief b_k that he updates via (1). This is not a harmless technical convenience: the defender's utility in Eq. (4) is multiplied by the indicator 1_{b_k(θ)≤η}, and the optimal switching policy in (9) is built around keeping the attacker's belief low. An attacker who optimally conditions on his belief, as perfect Bayesian rationality normally requires, would recognize that deception is active, change his actions, and thereby alter the belief trajectory and the defender's value computed in Figs. 2–3. The paper therefore computes a constrained best response against a non-adaptive attacker rather than a Stackelberg equilibrium. Please either allow belief-dependent strategies in the lower layer or explicitly frame the model as one with a naive attacker and adjust the claims accordingly.
[Eqs. (1), (10) and Algorithm 1] The belief update in (1) and (10) is written for a fixed game mode θ over the horizon, but the strategic layer switches modes at stages selected by the dynamic program. After a switch to θ_{k+1}, the transition is generated by T^{θ_{k+1}}, yet the update in (10) computes the posterior over θ using a likelihood that does not account for the switching policy or for the fact that the active mode may have changed between stages. The extended state x̄_{k+1} contains θ_{k+1} as the new mode, but the belief b_{k+1} from (10) is a posterior over the mode that generated the last transition, which is the pre-switch mode. The paper never re-conditions b_{k+1} on the switch, so b_{k+1} is not the attacker's belief about the active mode θ_{k+1} if the attacker is aware of the strategic-layer policy. This inconsistency undermines the optimality claim of the switching policy and needs to be resolved explicitly.
[Section IV and abstract] The abstract and Section I claim that strategically timed deceptions reduce the risk of asset compromise by 25–50% compared to baselines without switching, but Section IV does not report any specific percentage reduction. Figures 2 and 3 plot values and total rewards, but no table gives the relative improvement, and the belief threshold η in Eq. (4) is listed in Algorithm 1 as an input but is never assigned a numerical value in Section IV-A. Without the value of η and exact numerical comparisons, the central quantitative claim is not reproducible or testable. Please add a table with the percentage improvements and state all parameter values used in the experiments.

minor comments (3)

[Section IV-B] There is a typo in the 'Fixed timing' bullet: 'timeing' should be 'timing'.
[Eq. (1) and Section II-A] The belief in (1) is written as b_k(θ|s_k), conditioning only on the current state, while the right-hand side includes the policy terms π_a^k(a_a|s_k) and π_d^k(a_d|s_k,θ). This notation loses the dependence on the history of actions and observations; please clarify whether the belief is intended to be a Markovian belief or a full-information posterior, and define the history space explicitly.
[Definition 3] In (C2), sequential rationality is stated for 'all given game mode θ' and an initial state, but the TP π* is a single profile over all modes. Please clarify how the equilibrium definition handles the fact that the attacker's belief and the defender's strategy must be simultaneously optimal for every possible mode when only one mode is active at a time.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the switching policy and baselines are computed from stated model primitives, and the improvement is a model consequence, not a fitted or self-cited conclusion.

full rationale

The paper's derivation chain is not circular. The lower tactical layer defines a one-sided information Markov game and an epsilon-PBNE with explicit belief consistency (Eq. (1)) and sequential rationality (Eqs. (6)-(7)). The upper strategic layer then solves a dynamic program (Eqs. (8)-(9)) whose value recursion compares staying versus switching using the tactical equilibrium and the stated utility U^theta from Eq. (4). The baselines in Section IV-B (no switch, fixed timing, uniform interval) are independent strategies; the proposed strategy is obtained by optimizing the same objective, not by fitting a parameter to those baselines. The numerical 25-50% improvement is a computed consequence of the model's reward and transition structure, not an input disguised as a result. The threshold eta in Eq. (4) is an unstated numerical parameter, which is a reproducibility weakness, but it is not a fitted value that is later renamed a prediction. The citations [11], [12] are used for a standard dynamic-programming solution method; they are not load-bearing for the central timing claim. The restriction to Markov state-only attacker strategies (Section II-A) is a modeling assumption that may limit the equilibrium's strength, but it is an explicit assumption, not a circular step. No step reduces by construction to its own inputs.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework itself has no fitted constants, but the numerical demonstration requires several hand-chosen values. The most concerning is the belief threshold eta in Eq. (4), which defines the defender's payoff and is never specified, and the hand-set alpha, delta, beta and reward tables that entirely determine the reported 25-50% improvement. The modeling also relies on nonstandard restrictions: the attacker's Bayesian belief is updated but not used in his Markov strategy.

free parameters (5)

belief threshold eta
Appears in Eq. (4) as the cutoff for the defender's utility U^theta(x_k, ...) = r^theta_d(...) * 1_{b_k(theta) <= eta}; no value is given in the numerical setup or experiments, yet it directly controls the reward used for both stay and switch values.
attacker ability alpha, defender ability delta, state-impact beta = alpha=0.8, delta=0.5, beta=infinity
Chosen by hand in Table II to define all transition probabilities; the reported utility improvement depends on these values.
immediate reward table r^theta_d = 10.0, 5.0, 1.0, 0.0 depending on action match (Table I)
Hand-specified reward values that define the tactical and strategic payoff structure; no justification or sensitivity analysis is provided.
terminal rewards U_T = 100, 50, 10, 0, -100 for s1..s5
Hand-set terminal payoffs in Table III that anchor the value of reaching each state.
attacker prior belief b0
The belief update (1) requires a prior; the numerical experiments do not report the prior, which affects the utility threshold in Eq. (4).

assumptions (4)

domain assumption A deception technique cannot be reused once the attacker recognizes it
Stated in Section I and II-B; this justifies the no-reuse and budgeted switching structure.
domain assumption The attacker updates beliefs via Bayes' rule (Eq. 1) and has a known initial prior b0
The belief dynamics are assumed known to the modeler; the paper does not justify this as the true attacker behavior.
ad hoc to paper The attacker is restricted to state-only Markov strategies despite Bayesian beliefs
Section II-A defines pi_a: S -> Delta(A_a), so the attacker cannot condition actions on belief b_k; this is not a standard PBNE restriction and weakens the equilibrium concept.
domain assumption Finite attacker lifetime K and budget M are known to the defender
The optimal stopping formulation in Section II-B takes K and M as inputs; in practice these may be unknown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of When to Deceive: A Cross-Layer Stackelberg Game Framework for Strategic Timing of Cyber Deception." pith.science (2026). https://pith.science/paper/MZXRJIYU

@misc{pith2026250521244,
  author       = {Pith},
  title        = {Pith review of: When to Deceive: A Cross-Layer Stackelberg Game Framework for Strategic Timing of Cyber Deception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MZXRJIYU}},
  note         = {Machine review of arXiv:2505.21244}
}

read the original abstract

Cyber deception is an emerging proactive defense strategy to counter increasingly sophisticated attacks such as Advanced Persistent Threats (APTs) by misleading and distracting attackers from critical assets. However, since deception techniques incur costs and may lose effectiveness over time, defenders must strategically time and select them to adapt to the dynamic system and the attacker's responses. In this study, we propose a Stackelberg game-based framework to design strategic timing for cyber deception: the lower tactical layer (follower) captures the evolving attacker-defender dynamics under a given deception through a one-sided information Markov game, while the upper strategic layer (leader) employs a stopping-time decision process to optimize the timing and selection of deception techniques. We also introduce a computational algorithm that integrates dynamic programming and belief-state updates to account for the attacker's adaptive behavior and limited deception resources. Numerical experiments validate the framework, showing that strategically timed deceptions can enhance the defender's expected utility and reduce the risk of asset compromise compared to baseline strategies.

Figures

Figures reproduced from arXiv: 2505.21244 by the authors.

**Figure 2.** The results show the value at the initial stage for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** The results for the total rewards (sum of immediate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Guarding Against Malicious Biased Threats (GAMBiT) Experiments: Revealing Cognitive Bias in Human-Subjects Red-Team Cyber Range Operations
cs.CR 2025-08 conditional novelty 6.0 of 10

Three multi-modal datasets capture 59 skilled attackers' full operational traces (keystrokes, shell history, PCAP, surveys) in a simulated enterprise network, with labels designed to reveal cognitive biases.
Bi-Level Game-Theoretic Planning of Cyber Deception for Cognitive Arbitrage
cs.GT 2025-09 conditional novelty 5.0 of 10

A bi-level game-theoretic framework for timing and selecting cyber deception against biased attackers, with simulated evidence that optimal switching yields at least a 40% reward improvement.
A Multi-Resolution Dynamic Game Framework for Cross-Echelon Decision-Making in Cyber Warfare
cs.CR 2025-07 conditional novelty 5.0 of 10

A two-resolution game framework lets cyber defenders zoom between tactical game trees and strategic Markov game states to refine defense plans.

Reference graph

Works this paper leans on

12 extracted references · 7 canonical work pages · cited by 3 Pith papers

[1]

A survey on advanced persistent threats: Techniques, solutions, challenges, and research opportunities,

A. Alshamrani, S. Myneni, A. Chowdhary, and D. Huang, “A survey on advanced persistent threats: Techniques, solutions, challenges, and research opportunities,”IEEE Communications Surveys & Tutorials, vol. 21, no. 2, pp. 1851–1877, 2019

2019
[2]

Mitre att&ck: Design and philosophy,

B. E. Strom, A. Applebaum, D. P. Miller, K. C. Nickels, A. G. Pennington, and C. B. Thomas, “Mitre att&ck: Design and philosophy,” inTechnical report. The MITRE Corporation, 2018

2018
[3]

Symbiotic game and foundation models for cy- ber deception operations in strategic cyber warfare,

T. Li and Q. Zhu, “Symbiotic game and foundation models for cy- ber deception operations in strategic cyber warfare,”arXiv preprint arXiv:2403.10570, 2024

arXiv 2024
[4]

Jajodia, A

S. Jajodia, A. K. Ghosh, V . Swarup, C. Wang, and X. S. Wang,Mov- ing target defense: creating asymmetric uncertainty for cyber threats. Springer Science & Business Media, 2011, vol. 54

work page 2011
[5]

A com- prehensive survey on cyber deception techniques to improve honeypot performance,

A. Javadpour, F. Ja’fari, T. Taleb, M. Shojafar, and C. Benza ¨ıd, “A com- prehensive survey on cyber deception techniques to improve honeypot performance,”Computers & Security, p. 103792, 2024

work page 2024
[6]

Behavioral cyber deception: A game and prospect theoretic approach,

S. Nan, S. Brahma, C. A. Kamhoua, and N. O. Leslie, “Behavioral cyber deception: A game and prospect theoretic approach,” in2019 IEEE Global Communications Conference (GLOBECOM). IEEE, 2019, pp. 1–6

work page 2019
[7]

Honeypot-based cyber deception against malicious reconnais- sance via hypergame theory,

A. H. Anwar, M. Zhu, Z. Wan, J.-H. Cho, C. A. Kamhoua, and M. P. Singh, “Honeypot-based cyber deception against malicious reconnais- sance via hypergame theory,” inGLOBECOM 2022-2022 IEEE Global Communications Conference. IEEE, 2022, pp. 3393–3398

work page 2022
[8]

Learning about cyber deception through simulations: Predictions of human deci- sion making with deceptive signals in stackelberg security games,

E. A. Cranford, C. Gonzalez, S. Cooney, and M. Tambe, “Learning about cyber deception through simulations: Predictions of human deci- sion making with deceptive signals in stackelberg security games,” in Proceedings of the Annual Meeting of the Cognitive Science Society, vol. 40, 2018

work page 2018

Show all 12 references

[9]

Bas ¸ar and G

T. Bas ¸ar and G. J. Olsder,Dynamic noncooperative game theory. SIAM, 1998

1998
[10]

The tularosa study: An experimental design and implementation to quantify the effectiveness of cyber deception

K. Ferguson-Walter, T. Shade, A. Rogers, M. C. S. Trumbo, K. S. Nauer, K. M. Divis, A. Jones, A. Combs, and R. G. Abbott, “The tularosa study: An experimental design and implementation to quantify the effectiveness of cyber deception.” Sandia National Lab.(SNL-NM), Albuquerque...

2018
[11]

A dynamic games approach to proactive defense strategies against advanced persistent threats in cyber-physical systems,

L. Huang and Q. Zhu, “A dynamic games approach to proactive defense strategies against advanced persistent threats in cyber-physical systems,” Computers & Security, vol. 89, p. 101660, 2020

2020
[12]

Gazeta: Game-theoretic zero-trust authentication for defense against lateral movement in 5g iot networks,

Y . Ge and Q. Zhu, “Gazeta: Game-theoretic zero-trust authentication for defense against lateral movement in 5g iot networks,”IEEE Trans- actions on Information Forensics and Security, vol. 19, pp. 540–554, 2023

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A survey on advanced persistent threats: Techniques, solutions, challenges, and research opportunities,

A. Alshamrani, S. Myneni, A. Chowdhary, and D. Huang, “A survey on advanced persistent threats: Techniques, solutions, challenges, and research opportunities,”IEEE Communications Surveys & Tutorials, vol. 21, no. 2, pp. 1851–1877, 2019

2019

[2] [2]

Mitre att&ck: Design and philosophy,

B. E. Strom, A. Applebaum, D. P. Miller, K. C. Nickels, A. G. Pennington, and C. B. Thomas, “Mitre att&ck: Design and philosophy,” inTechnical report. The MITRE Corporation, 2018

2018

[3] [3]

Symbiotic game and foundation models for cy- ber deception operations in strategic cyber warfare,

T. Li and Q. Zhu, “Symbiotic game and foundation models for cy- ber deception operations in strategic cyber warfare,”arXiv preprint arXiv:2403.10570, 2024

arXiv 2024

[4] [4]

Jajodia, A

S. Jajodia, A. K. Ghosh, V . Swarup, C. Wang, and X. S. Wang,Mov- ing target defense: creating asymmetric uncertainty for cyber threats. Springer Science & Business Media, 2011, vol. 54

work page 2011

[5] [5]

A com- prehensive survey on cyber deception techniques to improve honeypot performance,

A. Javadpour, F. Ja’fari, T. Taleb, M. Shojafar, and C. Benza ¨ıd, “A com- prehensive survey on cyber deception techniques to improve honeypot performance,”Computers & Security, p. 103792, 2024

work page 2024

[6] [6]

Behavioral cyber deception: A game and prospect theoretic approach,

S. Nan, S. Brahma, C. A. Kamhoua, and N. O. Leslie, “Behavioral cyber deception: A game and prospect theoretic approach,” in2019 IEEE Global Communications Conference (GLOBECOM). IEEE, 2019, pp. 1–6

work page 2019

[7] [7]

Honeypot-based cyber deception against malicious reconnais- sance via hypergame theory,

A. H. Anwar, M. Zhu, Z. Wan, J.-H. Cho, C. A. Kamhoua, and M. P. Singh, “Honeypot-based cyber deception against malicious reconnais- sance via hypergame theory,” inGLOBECOM 2022-2022 IEEE Global Communications Conference. IEEE, 2022, pp. 3393–3398

work page 2022

[8] [8]

Learning about cyber deception through simulations: Predictions of human deci- sion making with deceptive signals in stackelberg security games,

E. A. Cranford, C. Gonzalez, S. Cooney, and M. Tambe, “Learning about cyber deception through simulations: Predictions of human deci- sion making with deceptive signals in stackelberg security games,” in Proceedings of the Annual Meeting of the Cognitive Science Society, vol. 40, 2018

work page 2018