{"id":"bfedf172-e112-4370-97fb-61e6bbdf216c","arxiv_id":"1908.07466","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proposes a joint blockchain access control and double-dueling-deep-Q-network computation offloading scheme for IoT edge-cloud networks and claims lower long-term system cost than edge-only, cloud-only, and regular DQN baselines.","lead":"This paper combines blockchain-based access control with a deep reinforcement learning scheme to decide whether Internet of Things devices offload computation to an edge server or a remote cloud. The authors report lower latency and energy costs than baseline offloading policies, based on a private Ethereum experiment and numerical simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The (P1) objective is independent of the offloading decision vector A, so the DRL placement policy cannot be the source of the claimed cost savings; the Bellman equations in Eq. (7)-(10) also use min where the reward r=-C requires max.","rationale":"The reader's weakest-assumption is that the DRL state omits task size, channel gain, and transmit power, making the MDP non-Markovian. I find an even earlier structural defect: the optimization objective in (P1) does not depend on the offloading decision variables at all, because Eq. (5)-(6) sum edge and cloud costs instead of selecting one via α. Since each task is constrained to run on exactly one platform, the objective should be α_e C_e + α_c C_c; as written, every task incurs both edge and cloud latency/energy, so the binary decision disappears from the cost. This makes the claimed comparison of ADRLO against EO/CO baselines meaningless: the baseline difference can only arise from the f and w allocation, not from where tasks are offloaded. The ill-posed treatment of f^c, which appears in the cost but not as a constrained decision variable (it can grow without bound), compounds the problem. I also agree with the reader's state-insufficiency concern and note the Bellman min/max inconsistency, which is internal to the DRL derivation rather than a matter of consensus. The blockchain access-control experiment is independent, plausible, and shows real deployment effort; I credit it as engineering evidence. However, it does not address the offloading formulation. Because the central offloading claim rests on a malformed objective and a mismatched RL update, the REJECT verdict remains appropriate; my check would settle the issue by exposing the α-independence directly.","tokens_in":18699,"tokens_out":4477,"duration_ms":45623,"concrete_test":"Take the N=1 instance and fix D_1, X_1, τ_1, w_1, f^e_1, f^c_1, h_1, p_1. Substitute α=(1,0) and then α=(0,1) into the objective as defined by Eq. (5)-(6); the numerical value of C_1 is unchanged. This directly proves that the objective is independent of A. A complementary simulation check: re-run the Section VI-C experiments with A fixed to all-edge and then to all-cloud while optimizing f and w; if the two total-cost curves are identical, the offloading decision contributes nothing to the reported ADRLO gains. Independently, evaluate Eq. (8) with max instead of min on the stated reward and compare convergence; the min form cannot converge to the Q* for r=-C.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ADRLO learns an optimal offloading policy and achieves the lowest total cost (Section VI-C). This is unsupported by the model as written. In Section IV-B2, the per-device cost is defined as C_n = β_t T_n + β_e E_n with T_n = T_e_n + T_c_n and E_n = E_e_n + E_c_n (Eq. (5)-(6)). Each task is constrained to be executed at exactly one platform (C1-C2, α_e_n + α_c_n = 1), but the α variables do not appear in T_n or E_n: the objective in (P1) is identical for α_e=1 and α_c=1. Minimizing over A therefore cannot change the cost; if α is removed, (P1) is a resource-allocation problem only in f^e and w. Moreover, the cloud allocation f^c_n appears in T_c_n and E_c_n but is neither a decision variable nor constrained; if it were optimized, the cost would be driven downward by letting f^c_n → ∞. The DRL formulation inherits this flaw: the reward r(s,a) = -C(s,a) is independent of α, so no Q-learning agent can learn a placement policy from it. The state definition s={tc,ec,bw} omits D_n, h_n, p_n, which jointly determine the reward, so the MDP is not Markov. Eq. (7)-(10) and Algorithm 2 line 12 use min(Q) and argmin(Q), whereas with reward r=-C the Bellman optimality equation requires max(Q) and argmax(Q); min is correct only for a cost-to-go with r=C, not for the stated reward. The access-control demonstration (Section VI-B) is a plausible engineering validation, but it does not repair the offloading model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a secure computation-offloading framework for blockchain-based IoT networks, combining a smart-contract access-control mechanism with a deep reinforcement learning (DRL) offloading algorithm. The access-control part is implemented on a private Ethereum network and validated with two smartphones. The offloading part formulates a joint optimization problem (P1) over offloading decisions, edge computation resource, and radio bandwidth, and solves it with a double-dueling DQN algorithm (ADRLO). Numerical simulations compare ADRLO with regular DQN, edge-only, and cloud-only baselines, and report lower total cost. The core claims are that the smart contract rejects unauthorized offloading requests and that ADRLO learns an optimal offloading policy achieving the lowest system cost.","tokens_in":19110,"tokens_out":4683,"duration_ms":47789,"significance":"If the central claims were sound, the integration of blockchain-based access control with DRL-based offloading would be a plausible contribution to secure IoT edge-cloud systems. The access-control deployment and the honest reporting of the experimental setup are useful engineering pieces. However, the offloading model and the DRL formulation contain structural errors that invalidate the claimed optimality results: the objective in (P1) does not depend on the offloading decision variables, the cloud resource variable is unconstrained, the DRL state is not a sufficient statistic for the MDP, and the Bellman update equations use minimization where maximization is required. These are not local presentation issues; they undermine the core contribution, and the numerical results in Section VI-C cannot support the stated conclusions.","major_comments":[{"comment":"The cost function is independent of the offloading decision vector A. Equations (5) and (6) define T_n = T^e_n + T^c_n and E_n = E^e_n + E^c_n for every device, so the objective in (P1) sums edge and cloud costs regardless of whether the task is actually offloaded to the edge or the cloud. Since the binary variables alpha appear neither in the objective nor in constraints (C1)-(C6), every feasible A yields exactly the same cost, and minimizing over A is meaningless. The model must condition the latency and energy on the offloading decision, e.g., T_n = alpha^n_e T^e_n + alpha^n_c T^c_n, or otherwise make the objective depend on A.","section":"Section IV-B3, Eqs. (5)-(6), (P1)"},{"comment":"The cloud resource variable f^c_n appears in T^c_n and E^c_n but is neither a decision variable nor bounded by any constraint. If f^c_n is treated as a variable, the cost can be driven to minus infinity by letting f^c_n go to infinity; if it is a fixed parameter, its value and feasible range are never specified. This makes (P1) ill-posed. The paper states that cloud resource allocation is 'ignored,' but the cost expressions still depend on f^c_n, so the model is incomplete.","section":"Section IV-B3 and Section V-B1"},{"comment":"The state s = {tc, ec, bw} is not a sufficient statistic for the offloading MDP. The reward r(s,a) = -C(s,a) depends on the task data size D_n, the channel gain h_n, the transmit power p_n, the workload X_n, and the latency/energy weights, none of which appear in the state. Without these quantities, the agent cannot evaluate the reward associated with a state-action pair or predict future states, so the process is not Markov and the Q-learning convergence arguments do not apply. At a minimum, the state must include D_n, h_n, p_n, X_n, and the weights, or the authors must prove that the three listed components determine the full transition and reward structure.","section":"Section V-B1, state definition"},{"comment":"The Q-learning update, the Bellman optimality equation, the DQN target, and the double-DQN target all use min over the next action, while the stated reward is r(s,a) = -C(s,a). With a reward to be maximized, the correct operator is max, and the greedy action selection in Algorithm 2 line 12 should be argmax, not argmin. The text in Section V-B2 even states pi*(s) = argmax Q*(s,a), so the equations are inconsistent with the surrounding prose. If the authors intended a cost-to-go with r = C, then the reward definition in Section V-B1 must be changed accordingly.","section":"Section V-B2, Eqs. (7)-(10) and Algorithm 2 line 12"},{"comment":"Because the objective in (P1) and the DRL reward are independent of the offloading decision A, the simulation results cannot demonstrate that ADRLO learns an optimal placement policy. The reported cost differences between ADRLO, DRLO, edge-only, and cloud-only baselines cannot be attributed to the learned offloading decisions under the model as written; they would need to be explained by the resource-allocation components, which are not the paper's claimed contribution. The evaluation therefore does not support the central conclusion.","section":"Section VI-C, Figs. 8-10"}],"minor_comments":[{"comment":"The abstract states that the optimization minimizes latency, energy, and smart-contract fee, but the cost function in Section IV-B3 contains only weighted latency and energy; no smart-contract fee appears in (P1). Please reconcile these statements.","section":"Abstract vs. Section IV-B3"},{"comment":"The reference numbering is inconsistent: Ethereum is cited as [37], but [37] is an IoT access-control paper by Lin et al., and reference [38] duplicates the content of [20]. Please recheck all citations and the bibliography.","section":"References [37] and [38]"},{"comment":"The text defines beta^n_t and beta^n_e with an index n, but the cost expression C_n = beta_t T_n + beta_e E_n drops the index. Please clarify whether the weights are per-device or global.","section":"Section IV-B3, cost weights"},{"comment":"The target y^dou_j uses both theta and theta-prime, but the roles of the two networks are not clearly described in the text or pseudocode. A brief explanation of which network selects and which evaluates would improve readability.","section":"Algorithm 2, line 19"},{"comment":"There are several typographical issues, including 'MEECO' for 'MECCO' in Section I and inconsistent use of 'DRL' and 'DLR' near Section V-B2. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The access-control experiment is a straightforward deployment of a smart contract and does not by itself constitute a novel theoretical contribution. The offloading model has a load-bearing flaw: the cost objective is independent of the binary offloading decision, making the DRL placement policy vacuous, and the cloud resource variable is unbounded. The state-space and Bellman equations are also incorrect. These issues are not reparable by minor edits; they require reformulating the optimization model and re-running all simulations, which is beyond the scope of a revision. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The access-control experiment is the only part that stands up: a real Ethereum deployment on AWS, two Android phones, a Solidity contract, and clear screenshots of authorized versus unauthorized offloading requests. It is a straightforward application of the authors' earlier design [15], but it is honest engineering and it works. Everything else in the paper is built on an offloading model that, as written, cannot support its claims.\n\nThe core problem is that the cost in (5)-(6) sums edge and cloud latency and energy for every task. The binary offloading variables alpha_e, alpha_c never enter T_n or E_n, so the objective in (P1) is identical for alpha_e=1 and alpha_c=1. Minimizing over the offloading decision vector A is vacuous. The cloud CPU allocation f^c_n appears in the cost but is neither a decision variable nor bounded; the paper says cloud resource allocation is ignored, but then f^c_n shows up in T_c and E_c anyway. The DRL formulation inherits both problems: the reward r(s,a)=-C(s,a) is independent of alpha, so no agent can learn a placement policy, and the state s={tc,ec,bw} omits task size, channel gain, and transmit power, which determine the reward. The MDP is not Markov. On top of that, the Bellman equations in (7)-(10) and Algorithm 2 use min/argmin where the reward sign requires max/argmax. That is a direct sign error in the learning rule.\n\nThe evaluation compares ADRLO against fixed edge-only and cloud-only policies and a plain DQN. Because the objective does not depend on offloading decisions, the claimed cost savings cannot be attributed to better placement; they would have to come from resource allocation, but the DRL action includes alpha and the state does not include the parameters that matter. No code, no data, no error bars.\n\nI would not send this to a serious referee in its current form. The offloading half needs to be reworked from the cost formulation up. The access-control demo could be salvaged as a short applied paper, but the joint MECCO claim as it stands is not supported. Desk reject.","headline":"The blockchain access-control demo is real, but the offloading optimization is broken as written: the objective does not depend on the offloading decisions, so the paper's central performance claim is unsupported.","tokens_in":19652,"tokens_out":3862,"would_cite":false,"duration_ms":34651,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that blockchain-based access control and a double-dueling deep Q-network can jointly make IoT computation offloading secure and cost-efficient.","keywords":["blockchain","computation offloading","deep reinforcement learning","double dueling DQN","mobile edge computing","access control","smart contract","IoT networks"],"falsifier":"Take a single-device instance with fixed data size, channel gain, transmit power, and weights, enumerate all offloading and resource actions to find the exact minimum of the cost function, and compare that action with what ADRLO selects under its three-component state; any mismatch over a range of task sizes would show the state is not sufficient and the claimed optimal policy is unsupported.","tokens_in":18468,"feed_emoji":"⛓️","tokens_out":7754,"duration_ms":66212,"temperature":0.7,"pith_summary":"This paper argues that security and computation offloading in IoT edge-cloud networks can be handled in one framework: a blockchain smart contract decides which mobile devices may offload, and an advanced deep reinforcement learning agent—combining double and dueling Q-networks (ADRLO)—decides for each authorized device whether to send a task to the edge server or the cloud and how much computing and bandwidth to allocate. The goal is to minimize the long-term weighted sum of latency and energy cost across all devices. The paper reports that ADRLO achieves the lowest total cost against a plain DRL scheme, an all-edge scheme, and an all-cloud scheme, and that the smart contract rejects unauthorized offloading requests. If correct, the result suggests that untrusted IoT devices can get secure, cost-efficient offloading without relying on a centralized authority.","feed_headline":"Deep RL plus blockchain cuts IoT offloading cost in tests","feed_subtitle":"Smart contracts block unauthorized devices while a double-dueling Q-network chooses edge or cloud per task","key_machinery":"The load-bearing mechanism is the ADRLO agent: a deep Q-network that combines double DQN—which uses one network to choose the best action and another to evaluate it, reducing overestimation—with a dueling architecture that splits $Q(s,a)$ into a state-value $V(s)$ and an action-advantage $A(a)$ before recombining them. The agent observes state $s=\\{t_c,e_c,b_w\\}$ (total offloading cost, available edge computation, available bandwidth), selects an action containing offloading decisions plus edge and bandwidth allocations, and receives reward $r(s,a)=-C(s,a)$. Around it, a smart contract with add-device, delete-device, policy-list, and penalty functions enforces access control so only registered devices can reach the offloading stage.","core_discovery":"On its own terms, the paper establishes a joint access-control and offloading design for a mobile edge-cloud computation offloading (MECCO) system. Authorized devices are certified by checking a public key and device ID against a policy list in a smart contract; unauthorized requests are denied and penalized. For authorized devices, the offloading problem is formulated as minimizing $\\sum_{n=1}^{N} C_n$ where $C_n = \\beta_t T_n + \\beta_e E_n$, over binary edge/cloud decisions, edge CPU allocation, and bandwidth allocation. The paper reports that its advanced double-dueling DQN algorithm (ADRLO) yields the lowest system cost under varying numbers of devices, task sizes, edge capacities, and total bandwidths, and that jointly including edge and bandwidth allocation is necessary to achieve those savings.","pith_inferences":["The authors do not put the task data size, channel gain, transmit power, or cost weights into the state, so a natural extension is to test whether adding them changes the learned policy or lowers measured cost.","The paper's access-control experiment penalizes an unauthorized request only with a warning message; whether a real deterrent such as a fee or reputation loss would reduce repeated attacks is left untested.","Blockchain transaction and mining overhead is not included in the offloading cost model, so the reported cost savings likely do not include the security tax a real deployment would pay."],"forward_implications":["Authorized mobile devices can dynamically choose between edge and cloud execution depending on task size, so small tasks stay at the edge while large tasks go to the cloud.","The deep RL approach removes the need for prior knowledge of system statistics, because the agent learns an offloading policy from trial-and-error interactions.","Edge computation capacity and radio bandwidth should be allocated jointly with offloading decisions; dropping either allocation degrades total cost, as shown by the ablations in Fig. 10.","Blockchain-based access control replaces a single trusted authority with a distributed, auditable ledger of offloading transactions and device permissions."],"supporting_citations":[{"why":"Supplies the blockchain-based access control architecture and prior efficiency comparison that the MECCO security design builds on.","marker":"[15]"},{"why":"Provides the multiuser joint offloading and resource optimization task model, including the program profiler assumption for $D_n$ and $X_n$.","marker":"[23]"},{"why":"Supplies the mixed fog/cloud offloading model with the transmission rate formula and the latency and energy equations (1)-(4) that the paper adapts.","marker":"[25]"},{"why":"Provides the Q-learning update, Bellman optimality equation, and convergence claim that the DRL solution relies on.","marker":"[27]"},{"why":"Introduces the deep Q-network with experience replay that forms the base of the ADRLO algorithm.","marker":"[29]"},{"why":"Supplies the double DQN technique that ADRLO uses to reduce overestimation of action values.","marker":"[45]"},{"why":"Supplies the dueling network architecture that ADRLO uses to split state-value and advantage estimates.","marker":"[46]"}],"fun_headline_variants":["Blockchain + deep RL secure and cut IoT offloading costs","Double-dueling Q-net optimizes secure IoT offloading","Smart contracts guard edge-cloud offloading in IoT, RL saves cost","Deep RL and blockchain minimize delay-energy-fee in IoT offload"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three-part state—current cost, remaining edge computation, and remaining bandwidth—contains enough information for the agent to learn an optimal offloading policy, even though the task data size, channel gain, transmit power, and cost weights that determine the reward are not part of the state.","fun_headline_variants_meta":{"raw":{"variants":["Blockchain + deep RL secure and cut IoT offloading costs","Double-dueling Q-net optimizes secure IoT offloading","Smart contracts guard edge-cloud offloading in IoT, RL saves cost","Deep RL and blockchain minimize delay-energy-fee in IoT offload"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2908,"prompt_tokens":926,"completion_tokens":1982,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1910}},"tokens_in":542,"tokens_out":1982,"duration_ms":15227,"temperature":1.0,"reasoning_tokens":1910,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:13:34.806486+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single-device instance with fixed data size, channel gain, transmit power, and weights, enumerate all offloading and resource actions to find the exact minimum of the cost function, and compare that action with what ADRLO selects under its three-component state; any mismatch over a range of task sizes would show the state is not sufficient and the claimed optimal policy is unsupported.","supporting_citations":[{"cited_title":"Nguyen, Pubudu N","cited_arxiv_id":null,"evidence_quote":"Supplies the blockchain-based access control architecture and prior efficiency comparison that the MECCO security design builds on."},{"cited_title":"Lyu et al., ”Multiuser Joint Task Offloading and Resource Optimization in Proximate Clouds,” IEEE Transactions on Vehicular Technology, vol","cited_arxiv_id":null,"evidence_quote":"Provides the multiuser joint offloading and resource optimization task model, including the program profiler assumption for $D_n$ and $X_n$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mixed fog/cloud offloading model with the transmission rate formula and the latency and energy equations (1)-(4) that the paper adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Q-learning update, Bellman optimality equation, and convergence claim that the DRL solution relies on."},{"cited_title":"Kavukcuoglu, D","cited_arxiv_id":null,"evidence_quote":"Introduces the deep Q-network with experience replay that forms the base of the ADRLO algorithm."},{"cited_title":"Van Hasselt et al., Deep reinforcement learning with double q-learning,” in Conf","cited_arxiv_id":null,"evidence_quote":"Supplies the double DQN technique that ADRLO uses to reduce overestimation of action values."}],"review_version":1}