{"id":"c9056863-7439-486d-a8bb-d73015200714","arxiv_id":"1908.07467","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A DQN-based offloading policy for mobile blockchain miners reduces simulated latency and energy costs and increases a formula-based privacy score compared with fixed offload/no-offload baselines.","lead":"This paper trains Q-learning and deep Q-network agents to decide whether mobile blockchain miners should compute IoT data tasks locally or upload them to an edge server. The authors report that the deep RL policy cuts simulated latency and battery use while raising a self-defined privacy score.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The privacy claim rests on an unvalidated proxy metric that is also the training objective; without an independent attacker-based check, the 'significantly enhance user privacy' conclusion is not supported.","rationale":"I read the paper in good faith. The proposed MEC-blockchain offloading architecture is coherent, and the experiments in Fig. 6 and the simulations in Tables II–III may well support the latency and energy claims if the cost model is corrected. However, the paper's distinguishing contribution is privacy preservation, and that claim depends entirely on the proxy in Eqs. (8)–(10). The reader's weakest-assumption analysis identifies exactly this point, and I agree with it. The issue is not merely that the metric comes from prior work; it is that the metric is both the training signal and the evaluation criterion, and it has no demonstrated relationship to an actual adversary's inference success. Equations (8) and (9) are deterministic functions of the binary offloading action and the data sizes, so maximizing them can change offloading behavior without necessarily reducing real privacy leakage. Because the central claim is testable with an independent attacker model, the appropriate verdict remains CONDITIONAL: if the concrete test shows no privacy benefit under an external adversary, the privacy claim should be rejected or heavily qualified; if it shows a benefit, the current conditional acceptance is justified. I therefore recommend no change to the reader's verdict.","tokens_in":21933,"tokens_out":5000,"duration_ms":53977,"concrete_test":"Re-run the privacy evaluation of Section VII-C4 with an independent attacker instead of Eq. (10). Have the MEC server observe each miner's offloading decision x_t, data sizes D0_t and D1_t, and channel state g_t, then train a classifier to infer (i) the miner's ground-truth location (e.g., distance to the access point) and (ii) whether the sensing data D1_t corresponds to a new event (usage pattern). Compare attacker accuracy or AUC under DRLO, RLO, EO, and NO. If DRLO's higher Eq. (10) score does not translate into a meaningful reduction in attacker accuracy, or if its advantage disappears, the privacy-enhancement claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the privacy quantity P_nm^t defined in Eqs. (8)–(10) faithfully measures a miner's actual privacy against the curious MEC server and eavesdroppers described in Section IV-B. This same quantity is what the RLO and DRLO agents are trained to maximize, so the Fig. 10 comparison largely shows how well each policy optimizes its own objective rather than whether an adversary's inference is actually impeded. The formulas reinforce the concern: with D > 0 and x in {0,1}, Eq. (9) reduces to exactly the offloading decision under a bad channel, and Eq. (8) is simply |D0 - D1| or |D0| under a good channel. Neither expression is derived from an attacker's estimation error, location uncertainty, or usage-pattern inference, and no independent privacy validation is provided. If this proxy is not validated, the paper's distinct claim of privacy enhancement is unsupported even if the latency and energy comparisons are internally consistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a mobile edge computing (MEC) based blockchain architecture in which mobile users act as miners and offload both IoT data-processing tasks and mining tasks to a MEC server. The offloading decision is formulated as a Markov decision process whose objective combines user privacy, mining reward, and offloading latency/energy cost, and two learning algorithms are proposed: a tabular Q-learning scheme (RLO) and a deep Q-network scheme (DRLO). The paper reports experiments comparing local and edge computation, and simulations comparing DRLO/RLO with non-offloading, edge-offloading, and CMDP-based baselines in terms of latency, power, cost, and a privacy metric. The main claimed contributions are that DRLO significantly enhances user privacy and reduces energy consumption and computation latency at minimum offloading cost.","tokens_in":22098,"tokens_out":6481,"duration_ms":64041,"significance":"If the central claims were fully supported, the paper would address a genuinely underexplored problem: privacy-aware computation offloading in MEC-based mobile blockchain networks, where earlier work has focused mainly on offloading efficiency and mining economics. The proposed DQN formulation and the inclusion of both IoT data tasks and mining tasks are reasonable extensions of existing RL-based offloading work, and the paper provides a fairly complete set of comparisons against common baselines. However, the two most distinctive claims—privacy enhancement and the joint optimization of privacy, mining reward, and cost—currently rest on an unvalidated and circular privacy metric and on a reward function that is not dimensionally consistent. The latency and energy comparisons are internally plausible, but the main intellectual contributions need substantial additional support before the conclusions can be accepted.","major_comments":[{"comment":"The immediate reward in Eq. (17) is the sum of the privacy level P_t (a data size, in bits, from Eqs. (8)-(10)), the mining reward R_mining_n (in tokens, Eq. (13)), and the negative cost C_t (a weighted sum of energy in joules and latency in seconds from Eq. (15)). These quantities have different units and orders of magnitude, so the scalar objective being maximized by RLO and DRLO is not well defined; the tradeoff factors α1, α2, and λ do not have a common scale. Eq. (21), which is presented as a rewrite of Eq. (17), omits R_mining and still subtracts an energy-latency cost from a privacy term, so it does not resolve the inconsistency. The authors should normalize or rescale the components and demonstrate that the reported gains are not artifacts of the chosen units and scales.","section":"V-B, Eq. (17), and VII-C"},{"comment":"The privacy metric used for evaluation is the same function that appears in the RL reward, and it is not validated against any attacker model. With x_t=1, Eq. (8) reduces to D1 under a good channel; with x_t=0 it reduces to D0; and Eq. (9) is 1 exactly when the user offloads under a bad channel. Thus the metric largely encodes the offloading decision itself, rather than the inference error of the curious MEC server or an eavesdropper described in Section IV-B. The privacy comparison in Fig. 10 therefore largely shows how well each algorithm optimizes this particular formula, not whether user location or usage patterns are actually protected. An independent validation—for example, measuring an adversary's estimation error of D0/D1 or of user location under the learned policies—is needed to support the abstract's claim of \"significantly enhance user privacy.\"","section":"IV-B, Eqs. (8)-(10), and VII-C4, Fig. 10"},{"comment":"The model assumes in Sections III.A and IV.A that the MEC server has sufficient computation resources to serve all miners, and the optimization problem P1 contains no capacity constraint, only the binary and deadline constraints (16a)-(16b). Yet the discussion of Table II states that \"the computation capacity of the MEC server is only sufficient to provide resources for a certain number of miners\" and attributes part of the EO scheme's poor performance to capacity limits. This contradiction affects the multi-user claims: if capacity is unlimited, the EO cost should not grow because of resource scarcity; if it is limited, the problem formulation and algorithms are missing a resource-allocation component. The authors should either remove the capacity discussion or add the corresponding constraint and incorporate capacity into the optimization.","section":"III-A, IV-A, and VII-C3, Table II"},{"comment":"The mining reward term in the optimization objective cannot influence the learned offloading policy as formulated. R_mining_n in Eq. (13) depends only on p_n, H, s_n, and Y_n, while the action space in Section V-B is restricted to the offloading decision vector x_t; hash power and payment are treated as predetermined. Adding a constant with respect to the action to the reward does not change the argmax policy, so the joint optimization of mining profit and offloading is not actually implemented. The authors should either include mining-related actions in the MDP or explicitly state that the mining reward is a fixed offset and remove it from the objective that the RL agent optimizes.","section":"V-B, Eqs. (13) and (17); VI-B"},{"comment":"The experimental evaluation in Section VII.B compares only local execution with edge execution in terms of processing time and battery consumption; it does not run the RLO or DRLO algorithms or compare them with the baselines. The abstract, however, states that experiments and simulations show the proposed schemes enhance privacy and reduce energy and latency. The experiments should either be integrated with the algorithm evaluation or the claims should be limited to the simulations; otherwise, the experimental evidence does not support the paper's stated contribution.","section":"VII-B and Abstract"}],"minor_comments":[{"comment":"Algorithm 1, line 12, and Algorithm 2, line 10, contain a typographical error: \"D1_0\" should be \"D_t_0nm\" in the local-execution expression.","section":"Algorithms 1 and 2"},{"comment":"There is a typo in the text before Eq. (15): \"taks\" should be \"tasks.\"","section":"Eq. (15)"},{"comment":"In the final paragraph of Section VII.D, \"adative\" should be \"adaptive.\"","section":"Section VII.D"},{"comment":"Algorithm 2 does not implement a target network or a periodic target-network update, but Section VII.C.1 attributes the stability of DRLO to \"the iterative target network.\" Either implement the target network as described or remove that explanation.","section":"VI-B, Algorithm 2, and VII-C1"},{"comment":"The DQN architecture is selected by inspecting the same system-reward metric that the algorithm is trained to maximize, which risks overfitting the reported curve; a separate validation set or a description of the tuning protocol would strengthen the claim.","section":"VII-C1, Fig. 5"},{"comment":"No code or data availability statement is provided; releasing the simulation code and hyperparameter configurations would allow the community to reproduce the reported numbers, which would be particularly valuable given the number of tunable parameters in the reward and DQN.","section":"General reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is closely paired with arXiv:1908.07466 by the same group, and the simulation framework, parameter settings, and parts of the background overlap substantially with that companion work. The editor may wish to check the journal's dual-submission and self-plagiarism policies and ask the authors to clarify the relationship between the two manuscripts. The main technical concerns above are substantive but fixable in revision: the reward function needs rescaling, the privacy claim needs an independent attacker-based validation, and the capacity assumption needs to be made consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a competent engineering combination, not a conceptual breakthrough. The new part is putting privacy, mining reward, and offloading cost into one DQN objective for MEC-enabled mobile blockchain. The experiments (Android client, AWS Lambda edge) and the single/multi-user latency and power comparisons are real evidence and look internally consistent. The DRLO scheme beating RLO and the baselines on its own cost metric is plausible.\n\nThe soft spot is the privacy claim. The privacy level in Eqs. (8)-(10) is taken directly from [29] and never validated against an attacker model. That formula is also the part of the reward the agents are trained to maximize, so Fig. 10 mostly shows which policy better optimizes that formula, not whether a curious MEC server or eavesdropper actually learns less. The abstract's 'significantly enhance user privacy' is not supported by the evidence in the paper. The authors should either validate against a concrete inference attack or weaken the claim to 'privacy-aware offloading objective.'\n\nOther issues are smaller but real. The reward mixes bits, tokens, and weighted energy/latency without normalization; it functions as a heuristic, which is fine, but the optimization claim is loose. The paper assumes sufficient MEC capacity in Sec. III and then talks about capacity limits in the multi-user results. The DQN hyperparameters are selected by looking at the same system reward used for evaluation, which is circular. The closely related DRL blockchain offloading work [37] is cited but not positioned or benchmarked. No code or data are released.\n\nNone of this kills the latency and energy contribution. Those results can stand on their own. The citation pattern is honest: the borrowed pieces are attributed. The paper is worth a serious referee, because the formulation is reasonable and the experiments are a step beyond pure simulation. I would send it out, with a clear request that the privacy claim be either validated or reframed.","headline":"A workmanlike DRL offloading paper whose latency/energy results are plausible but whose privacy claim rests on a borrowed, unvalidated proxy that is also the training objective.","tokens_in":22670,"tokens_out":2622,"would_cite":false,"duration_ms":25830,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a deep Q-network offloading policy can simultaneously raise user privacy and cut energy and latency for mobile blockchain miners, beating local-only, edge-only, and prior RL and CMDP baselines.","keywords":["blockchain","mobile edge computing","task offloading","mining","privacy","deep reinforcement learning","deep Q-network","Q-learning"],"falsifier":"An adversary that infers a miner's location or usage pattern from the offloading history would settle the claim: if the DRLO policy does not reduce the adversary's inference accuracy compared with the baselines, or if the privacy score from Eq. (10) rises while a leakage measure such as mutual information between the offloading pattern and the true location stays flat or rises, then the central privacy enhancement claim fails.","tokens_in":21679,"feed_emoji":"🤖","tokens_out":8703,"duration_ms":77997,"temperature":0.7,"pith_summary":"This paper tries to show that a mobile blockchain miner with limited device resources can safely offload both IoT data-processing tasks and mining work to a nearby edge server, and that a deep reinforcement learning agent can find the offloading policy that best balances privacy, computation cost, and mining profit. The authors formulate the joint problem as a Markov decision process and solve it with Q-learning and, for large state spaces, a deep Q-network. If the claim is right, resource-constrained mobile users could participate in blockchain mining without giving away their location or usage patterns, while spending less energy and time than all-local or all-edge alternatives.","feed_headline":"Deep RL offloading lifts privacy, cuts blockchain costs","feed_subtitle":"A DQN-based policy for miners offloading to edge servers beats local-only, edge-only, and prior RL baselines.","key_machinery":"The load-bearing mechanism is the binary offloading decision $x^t_{nm}\\in\\{0,1\\}$ combined with a reward that couples privacy and cost: the usage-pattern privacy term $P^{u,t}_{nm}=|D^t_{0nm}-x^t_{nm}(D^t_{0nm}+D^t_{1nm})|\\,\\mathbb{I}(g^t_n\\ge\\zeta)$ and the location privacy term $P^{l,t}_{nm}=\\mathbb{I}[x^t_{nm}(D^t_{0nm}+D^t_{1nm})]\\,\\mathbb{I}(g^t_n<\\zeta)$ reward the miner for deviating from the obvious channel-based offloading pattern, while the cost term $C_t$ penalizes latency and energy. The learning machinery is Q-learning with $\\epsilon$-greedy exploration and Bellman optimality, and the deep version replaces the Q-table with a two-hidden-layer feedforward network trained by experience replay with loss $L_t(\\theta_t)=\\mathbb{E}[(r_t+\\gamma\\max Q(s_{t+1},a'|\\theta')-Q(s_t,a_t|\\theta_t))^2]$. This machinery carries the argument because every reported privacy and cost improvement is the same objective the agent is trained to optimize.","core_discovery":"The paper's central claim is that the joint problem of task offloading, privacy preservation, and mining profit in a MEC-based mobile blockchain can be cast as a Markov decision process and solved by a deep Q-network, yielding a policy $\\pi^*$ that at each slot selects the offloading action $x^t_{nm}\\in\\{0,1\\}$ maximizing $r_t(s,a)=P_t(s,a)+R^{\\text{mining}}_n-C_t(s,a)$. Here $P_t$ is the sum of usage-pattern privacy and location privacy from Eqs. (8)-(10), $R^{\\text{mining}}_n$ is the expected mining reward after orphaning risk, and $C_t$ is the weighted latency and energy cost. The authors report that DRLO, the DQN version, converges in about 2500 slots to a higher long-term reward than the Q-learning version RLO, achieves the lowest average latency, power, and offloading cost among DRLO, RLO, edge-only, and local-only schemes for single and multi-user scenarios, and improves the privacy metric by roughly 5-13% over the RL and CMDP baselines.","pith_inferences":["A direct stress test is to replace the privacy formula in Eqs. (8)-(10) with an information-theoretic leakage measure and retrain the same DRLO policy; if the advantage over baselines survives, the privacy claim does not hinge on the formula the agent was trained to maximize.","Because the paper assumes the MEC server has sufficient capacity, the same DQN architecture could be extended to a state that includes edge queue lengths and resource prices, turning the binary offloading decision into a congestion-aware policy.","The mining-reward term currently depends only on purchased hash power and orphaning probability; swapping in rewards from proof-of-stake or other consensus protocols would generalize the framework to non-PoW blockchains."],"forward_implications":["If DRLO is deployed, mobile blockchain miners can offload both IoT data tasks and mining puzzles to a nearby MEC server while maintaining a learned balance between privacy, latency, and energy, without knowing the system dynamics in advance.","DRLO's convergence after about 2,500 time slots means the offloading policy can be trained online in a dynamic channel and transaction environment, and it consistently outperforms the tabular Q-learning version RLO in long-term reward.","In a single-user scenario with 12 tasks and $\\beta=0.8$, DRLO cuts average offloading cost by about 18.7% relative to RLO and by 57% and 65% relative to edge-only and local-only offloading.","In multi-user scenarios with 10 miners, DRLO reduces average cost by about 12.7%, 52.5%, and 30.4% compared with RLO, edge-only, and local-only schemes, respectively.","The privacy level achieved by DRLO is about 5.2% and 12.7% higher than the RL-based and CMDP baselines at 10 kB transaction sizes, and about 5.5% and 13.4% higher at 100 kB."],"supporting_citations":[{"why":"supplies the usage-pattern and location privacy metric adopted in Eqs. (8)-(10) and the idea of learning privacy-aware offloading policies","marker":"[29]"},{"why":"provides the constrained-MDP privacy-aware offloading baseline and channel-state transition settings used for comparison","marker":"[28]"},{"why":"supplies the MEC-blockchain offloading model and the latency and energy formulations for edge computation","marker":"[22]"},{"why":"provides the DQN machinery (experience replay, target network, loss function) that the DRLO algorithm is built on","marker":"[17]"},{"why":"supplies blockchain mining simulation parameters such as block size, hash power, and mining reward, plus a related online DRL offloading approach","marker":"[37]"},{"why":"gives the Q-learning update, Bellman optimality equation, and convergence justification that RLO relies on","marker":"[15]"},{"why":"supports the leased-hash-power mining model that lets mobile users purchase hash power from edge-cloud providers","marker":"[30]"}],"fun_headline_variants":["Deep RL offloading enhances privacy, cuts costs","DQN task offloading improves privacy in mobile blockchain","Reinforcement learning optimizes blockchain offloading with privacy","Edge offloading via deep RL: privacy and efficiency","Deep Q-network minimizes blockchain offloading costs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the privacy level computed by Eqs. (8)-(10) faithfully measures how much location and usage-pattern information a real attacker could learn from the offloading history, since the paper's privacy improvements are measured with that same formula the RL agent is trained to maximize.","fun_headline_variants_meta":{"raw":{"variants":["Deep RL offloading enhances privacy, cuts costs","DQN task offloading improves privacy in mobile blockchain","Reinforcement learning optimizes blockchain offloading with privacy","Edge offloading via deep RL: privacy and efficiency","Deep Q-network minimizes blockchain offloading costs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1600,"prompt_tokens":1024,"completion_tokens":576,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":640,"tokens_out":576,"duration_ms":5589,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:12:58.195073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An adversary that infers a miner's location or usage pattern from the offloading history would settle the claim: if the DRLO policy does not reduce the adversary's inference accuracy compared with the baselines, or if the privacy score from Eq. (10) rises while a leakage measure such as mutual information between the offloading pattern and the true location stays flat or rises, then the central privacy enhancement claim fails.","supporting_citations":[{"cited_title":"Learning-Based Privacy-Aware Offloading for Healthcare IoT With Energy Harvesting,","cited_arxiv_id":null,"evidence_quote":"supplies the usage-pattern and location privacy metric adopted in Eqs. (8)-(10) and the idea of learning privacy-aware offloading policies"},{"cited_title":"Privacy-Aware Offloading in Mobile-Edge Computing,","cited_arxiv_id":null,"evidence_quote":"provides the constrained-MDP privacy-aware offloading baseline and channel-state transition settings used for comparison"},{"cited_title":"Computation Offloading and Content Caching in Wireless Blockchain Networks With Mobile Edge Computing,","cited_arxiv_id":null,"evidence_quote":"supplies the MEC-blockchain offloading model and the latency and energy formulations for edge computation"},{"cited_title":"Online Deep Reinforcement Learning for Computation Offloading in Blockchain- Empowered Mobile Edge Computing,","cited_arxiv_id":null,"evidence_quote":"supplies blockchain mining simulation parameters such as block size, hash power, and mining reward, plus a related online DRL offloading approach"},{"cited_title":"Security Implications of Blockchain Cloud with Analysis of Block Withholding Attack,","cited_arxiv_id":null,"evidence_quote":"supports the leased-hash-power mining model that lets mobile users purchase hash power from edge-cloud providers"}],"review_version":1}