{"id":"ab444b3a-bea3-4f6b-9e4f-5abef1b83ca1","arxiv_id":"2501.02350","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PM-Dedup reduces latency and cloud load in encrypted deduplication by pre-computing PoW challenges and share-indexes and running duplicate checks inside SGX enclaves on edge servers.","lead":"PM-Dedup moves part of the encrypted data deduplication process, including duplicate checks and proof-of-ownership, from the cloud into SGX enclaves on customer-side edge servers. It reports large latency reductions and lower cloud overhead for multi-branch organizations with high data redundancy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 81.8% latency reduction is computed using an assumed cloud/edge latency ratio (50.96) from an external study rather than a measured WAN delay in the testbed, so the central quantitative claim is only as strong as that unvalidated multiplier.","rationale":"The paper's central contribution is architectural: moving deduplication checks and PoW tasks to an edge server reduces latency. The evaluation's quantitative support for that contribution depends on the cloud/edge latency multiplier taken from Charyyev et al. Because the reported absolute latency values, and therefore the 81.8% figure, are generated by scaling cloud interactions by this assumed ratio rather than by measuring a realistic WAN delay, the headline claim is not robust unless the ratio is validated or swept. The paper is otherwise plausible: the system is implemented, real datasets are used, and the design is coherent. The SGX security assumption is explicitly acknowledged as a limitation and is common in TEE-based systems; it is less load-bearing for the headline quantitative claim than the latency multiplier. This concern confirms the reader's CONDITIONAL verdict rather than moving it: the paper should be accepted only if the latency claims are qualified or the evaluation is rerun with a measured ratio.","tokens_in":21218,"tokens_out":4935,"duration_ms":49654,"concrete_test":"Re-run Exp 1 with the edge-to-cloud latency set to the actual RTT measured from the edge server to a real cloud region, or sweep the ratio over the values in Figure 2 (1.5, 8.5, 17.79, 50.96), keeping all code paths unchanged. If the LAB overall-latency reduction versus DupLESS at ratio 8.5 falls materially below the reported 81.8% (e.g., below 50%), the headline claim must be reworded as conditional on the latency-ratio assumption. In the same run, implement the DupLESS baseline directly instead of modeling it, or identify explicitly which latency terms are modeled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.1 (Table 2) takes Charyyev et al.'s average cloud/edge latency ratio (50.96; median 17.79) and states it is used as the baseline for subsequent experiments. Section 6.2 reports no direct measurement of cloud-to-edge WAN latency; the testbed is a set of R430/i7 machines on one network. The comparison with DupLESS and SGXDedup is also explicitly model-based ('comparisons are based on the conceptual frameworks and methodologies outlined in these approaches'). Thus the absolute latencies in Figures 3–5, including the 81.8% 'up to' reduction, are produced by scaling every cloud interaction by an assumed multiplier. Since the distribution in Figure 2 is skewed (over 75% of users have ratios above 8.5, while the average is 50.96), using the average materially overstates the typical benefit. If the true ratio is 8.5 or the median 17.79, the savings from moving checks to the edge shrink roughly in proportion to the number of cloud round trips avoided. No sensitivity analysis over the ratio is provided, and no measured WAN latency is reported. This makes the central latency-reduction claim scenario-dependent rather than a validated property of the system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PM-Dedup, a source-based encrypted deduplication system that migrates part of the deduplication-checking and Proof-of-Ownership (PoW) workload from the cloud to SGX enclaves on organization edge servers. The cloud pre-computes challenges and responses, builds a share-index, and pushes these to edge enclaves; the edge server performs local and share-index checks and PoW verification, escalating only misses to the cloud. The authors evaluate a prototype on five real-world datasets and report latency reductions versus DupLESS and SGXDedup, with a headline 81.8% reduction in overall upload latency on the LAB dataset while preserving deduplication effectiveness and confidentiality.","tokens_in":21447,"tokens_out":5894,"duration_ms":56670,"significance":"If the quantitative claims are substantiated, PM-Dedup would be a useful practical step for secure deduplication in multi-branch organizations: the system is fully implemented (~3.2K lines of C), evaluated on five real-world traces, and the design is clearly motivated. The paper is also honest about its SGX security assumptions. The main contribution is credible as an architecture, but the central latency-reduction numbers currently rest on two fragile evaluation choices: an unvalidated cloud/edge latency multiplier and non-implemented comparison baselines. The share-index selection study is a genuine strength, especially the comparison of CMS, CMS+locality, and exact-frequency selection on real traces.","major_comments":[{"comment":"The overall latency results, including the headline 81.8% reduction versus DupLESS on the LAB dataset, are computed by scaling cloud interactions with the external average latency ratio of 50.96 from the Charyyev et al. study; no WAN latency is measured in the testbed, and no sensitivity analysis is provided. Since Figure 2 shows a strongly right-skewed distribution (over 75% of users have ratios above 8.5 while the average is 50.96), the reported savings are not robust to the choice of multiplier. The authors should either measure cloud-to-edge latency in their deployment or present Figures 3-5 for a range of representative ratios (e.g., 8.5, 17.79, 50.96) and report how the latency reductions vary.","section":"Section 6.1, Table 2 and Figure 2; Section 6.2, Figures 3-5"},{"comment":"The comparison with DupLESS and SGXDedup is explicitly described as \"based on the conceptual frameworks and methodologies outlined in these approaches,\" meaning the baseline latencies in Figures 3-5 are not measured from running these systems. The claimed reductions are therefore model-based estimates whose accuracy depends on unstated assumptions about the number of cloud round trips, PoW costs, and transfer costs in those baselines. The authors should either implement or run the baselines, or clearly present the comparison as an analytical model and validate the model's per-stage cost components against direct measurements.","section":"Section 6.2"},{"comment":"The pseudocode in Algorithm 4 returns false when the file hash d is not found in the file-level map F, whereas the surrounding text states that the edge server then proceeds to verify each chunk at chunk level. The VerifyChunks function is only invoked when file-level verification fails, not when the file is absent from F. This internal inconsistency means the dual-level PoW procedure cannot be executed as described. The algorithm should call VerifyChunks (or an equivalent chunk-level routine) in the else branch, and the seed sent to the client should be drawn from the challenge array rather than the response array.","section":"Section 4.2, Algorithm 4"},{"comment":"The conclusion claims PM-Dedup works \"without compromising security,\" but the security argument relies entirely on the assumption that the SGX enclave operates as intended, with the paper explicitly acknowledging that SGX is vulnerable to denial-of-service and side-channel attacks. No security evaluation, formal analysis, or experimental validation is provided for the threat model. This is especially important because source-based deduplication is known to be vulnerable to side-channel attacks from malicious clients. The authors should temper the security claim or add a concrete analysis of the residual risks and the conditions under which the confidentiality guarantees hold.","section":"Section 3.3 and Section 7"}],"minor_comments":[{"comment":"The variable \"idu\" used in Algorithm 4 is not defined in the data structure description; the text defines \"idc\" and \"cidc\" as the index counters. The algorithm should use the defined counters consistently.","section":"Section 4.2, Algorithm 4"},{"comment":"The sentence \"Figure 8 shows that the CMS approach yields the highest generation speed\" appears to refer to the generation-speed results, which are plotted in Figure 7; Figure 8 shows the MS deduplication ratio. The cross-reference should be corrected.","section":"Section 6.3"},{"comment":"There are several typos and heading issues: Section 1 is titled \"Introductions,\" Section 4 is titled \"Design ISSUES,\" and Section 6.2 contains \"lcoal-index\" in the paragraph before Figure 3. These should be fixed.","section":"Sections 1, 4, and 6.2"},{"comment":"The latency measurements are reported as single-point values from ten runs without confidence intervals or variances. Reporting the spread would make the comparisons more informative.","section":"Section 6.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the prototype is substantial, but the evaluation's reliance on an assumed latency multiplier and on non-implemented baselines makes the quantitative headline claims fragile. The Algorithm 4 inconsistency is a concrete correctness issue that should be fixed before publication. I would support acceptance after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PM-Dedup's architecture is the real contribution: moving dedup checks and PoW into SGX enclaves on branch edge servers, with a cloud-maintained share-index selected by CMS plus logical locality. That combination is new relative to SGXDedup and DupLESS, and the paper makes a sensible case that it cuts cloud round trips for backup workloads with high redundancy. The prototype is real: about 3.2K lines of C, FastCDC, SHA-256, AES-256, and they ran it on four public datasets plus their own LAB. The share-index selection evaluation (CMS vs frequency-based) is honest and shows the expected tradeoff: frequency-based has better dedup ratio but higher memory and compute; CMS plus locality degrades more gracefully over time. That part is solid.\n\nThe soft spot is the latency evaluation. The headline 81.8% reduction versus DupLESS is not measured against a real cloud. They never ran DupLESS or SGXDedup; they model those baselines from the papers. And they scale all cloud interaction latencies by a fixed multiplier (average 50.96) taken from Charyyev et al.'s latency study, rather than measuring WAN latency in their testbed. Since that distribution is skewed (median 17.79, and over 75% of users above 8.5), using the average inflates the benefit. No sensitivity analysis over the ratio. If the real ratio is 10 or 20, the savings shrink proportionally. This makes the central quantitative claim scenario-dependent, not a validated property of the system. That is the main reason the verdict should be conditional, not accept.\n\nOther issues are minor: no formal security analysis beyond the threat model; SGX side-channel/DoS attacks acknowledged but not mitigated; some pseudocode inconsistencies (Algorithm 3 uses SMK while the text says CSMK; Algorithm 4 uses fileRes as the challenge seed, likely a typo for fileChal). Free parameters like K, T, and update periods are not specified, so reproduction would require guessing.\n\nStill, the design is defensible and the paper is clearly written. The gap between claim and evidence is fixable: measure real WAN latency, run at least one actual baseline, and do a sensitivity sweep over the latency ratio. I would send it to review, but I would insist on those experiments before acceptance. For a reader, it is worth reading for the architecture and the share-index selection results, but not for the absolute latency numbers.","headline":"Architecture is new and plausible, but the headline latency reduction is an artifact of an assumed cloud/edge multiplier rather than a measured result.","tokens_in":22017,"tokens_out":4525,"would_cite":false,"duration_ms":37271,"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":"PM-Dedup cuts secure cloud deduplication latency by moving deduplication checks and proof-of-ownership from the cloud to SGX-protected edge servers, with up to 81.8% lower upload latency than DupLESS on a backup-heavy dataset.","keywords":["secure deduplication","encrypted deduplication","trusted execution environment","Intel SGX","edge computing","proof of ownership","share index","cloud storage"],"falsifier":"Deploy the prototype with real WAN links between client, edge, and cloud and measure upload latency for the LAB workload; if the measured cloud-to-edge latency ratio is close to 1, the latency reduction versus DupLESS should drop well below 81.8%, showing how much of the result depends on the assumed latency model.","tokens_in":20993,"feed_emoji":"☁️","tokens_out":3970,"duration_ms":36271,"temperature":0.7,"pith_summary":"PM-Dedup tries to solve a three-way tension in encrypted cloud storage: deduplication requires identical data to produce identical ciphertexts, but the standard source-based approach that saves bandwidth adds latency and cloud-side computation because every upload requires a deduplication check and a proof-of-ownership (PoW) round trip. The paper's proposal is to move part of that work from the cloud to an organization's edge servers, inside Intel SGX enclaves, using a cloud-selected share-index and precomputed challenge-response pairs. If the central claim is right, clients in redundant, branch-office workloads will see much lower upload latency—up to 81.8% less than the DupLESS baseline on the LAB dataset—without giving up the bandwidth savings of source-based deduplication or the confidentiality of client-side encryption.","feed_headline":"Edge servers cut secure cloud dedup latency by up to 82%","feed_subtitle":"PM-Dedup runs deduplication checks and proof-of-ownership inside SGX enclaves at the edge, easing the cloud bottleneck.","key_machinery":"The load-bearing mechanism is a tiered deduplication checkpoint at the edge, guarded by SGX. The cloud maintains a full-index of encrypted-chunk fingerprints and periodically selects a share-index using Count-Min Sketch, a space-efficient frequency estimator, optionally combined with logical-locality scoring from file recipes; this share-index and a set of precomputed challenge-response pairs for file- and chunk-level Proof of Ownership are stored inside the edge server's SGX enclave. When a client uploads, the edge server first checks a local LRU index, then the share-index, then escalates to the cloud only on a miss. Precomputed challenges let the edge verify ownership without real-time cryptographic computation, which is what makes the edge offload feasible given the enclave's limited compute.","core_discovery":"The central claim is that most of the latency and cloud overhead in secure source-based deduplication comes from cloud round trips for duplicate checking and PoW verification, and that these round trips can be largely eliminated by relocating the check to a client-side edge server while keeping the security properties intact. PM-Dedup does this by having the cloud precompute PoW challenges and responses for frequently accessed files and chunks, select a compact share-index of high-value fingerprints, and ship both to an SGX enclave on the edge server. The client then proves ownership and checks duplicates against the edge server first, using a local LRU index, then the share-index, then the cloud only as fallback. The paper reports that on backup-style datasets with high redundancy this design cuts overall upload latency by up to 81.8% versus DupLESS, and by smaller but positive margins on lower-redundancy datasets, while preserving deduplication effectiveness.","pith_inferences":["The 81.8% figure is tied to the assumed cloud-to-edge latency ratio, with an average ratio of about 51 from a public measurement study; in deployments where the cloud is only a few times slower than the edge, the latency savings will shrink and may not justify the extra edge hop.","The scheme inherits SGX's security boundary, so a practical SGX side-channel or denial-of-service attack would undercut the confidentiality argument that the migration relies on; the paper explicitly acknowledges this assumption.","The share-index is a heuristic, so its hit ratio depends on workload stability; workloads dominated by brand-new data could see negative returns from the extra edge checkpoint.","A natural extension would be adaptive share-index updates based on measured hit ratios, and a real-WAN evaluation across multiple branches to validate the latency model."],"forward_implications":["In high-redundancy backup workloads, most deduplication checks and PoW verifications complete at the edge, so cloud round trips stop being the latency bottleneck.","Cloud-side computational load from duplicate checks, challenge generation, and response verification is reduced because the cloud only handles share-index misses and periodic precomputation.","The same security model as server-aided message-locked encryption with PoW is preserved: clients still encrypt with key-server-derived keys and must prove ownership before deduplicating.","Because only unique chunks are uploaded, bandwidth savings are retained while latency drops, which is the combination target-based deduplication cannot offer.","The share-index plus locality selection keeps deduplication effectiveness close to exact frequency-based selection while using less memory and generation time."],"supporting_citations":[{"why":"DupLESS is the primary server-aided MLE baseline that PM-Dedup compares against for latency and security.","marker":"[17]"},{"why":"SGXDedup is the SGX-accelerated encrypted deduplication baseline that PM-Dedup extends and compares with.","marker":"[26]"},{"why":"Charyyev et al. supplies the real-world cloud-to-edge latency ratios used to scale cloud latency in the evaluation.","marker":"[4]"},{"why":"Halevi et al. provides the foundational proof-of-ownership protocol that PM-Dedup's challenge-response PoW builds on.","marker":"[10]"},{"why":"Di Pietro and Sorniotti supplies the efficiency-oriented PoW construction that motivates PM-Dedup's lightweight dual-level PoW.","marker":"[7]"},{"why":"Cormode and Muthukrishnan provides the Count-Min Sketch used for the lightweight share-index selection.","marker":"[6]"},{"why":"Bellare et al. defines message-locked encryption, the basis for deduplicating encrypted data with content-derived keys.","marker":"[2]"},{"why":"FastCDC is the content-defined chunking method used to split files into deduplication units in the client implementation.","marker":"[38]"}],"fun_headline_variants":["Edge TEEs cut cloud dedup latency up to 82%","PM-Dedup moves dedup checks to edge, cutting latency 82%","Secure dedup latency slashed by shifting checks to edge","Partial migration to edge cuts dedup upload time 82%","Edge enclaves reduce cloud dedup latency by up to 82%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the cloud-to-edge latency ratio measured in a public study, with an average of 50.96 and median of 17.79, applies to the target deployment and uses that ratio to scale cloud latency instead of measuring real WAN latency; if the true ratio is much lower, the reported latency gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["Edge TEEs cut cloud dedup latency up to 82%","PM-Dedup moves dedup checks to edge, cutting latency 82%","Secure dedup latency slashed by shifting checks to edge","Partial migration to edge cuts dedup upload time 82%","Edge enclaves reduce cloud dedup latency by up to 82%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2812,"prompt_tokens":1043,"completion_tokens":1769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":1690}},"tokens_in":659,"tokens_out":1769,"duration_ms":12975,"temperature":1.0,"reasoning_tokens":1690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:20.999604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the prototype with real WAN links between client, edge, and cloud and measure upload latency for the LAB workload; if the measured cloud-to-edge latency ratio is close to 1, the latency reduction versus DupLESS should drop well below 81.8%, showing how much of the result depends on the assumed latency model.","supporting_citations":[{"cited_title":"{DupLESS}:{Server-Aided} encryption for dedu- plicated storage","cited_arxiv_id":null,"evidence_quote":"DupLESS is the primary server-aided MLE baseline that PM-Dedup compares against for latency and security."},{"cited_title":"Accelerating encrypted dedupli- cation via {SGX}","cited_arxiv_id":null,"evidence_quote":"SGXDedup is the SGX-accelerated encrypted deduplication baseline that PM-Dedup extends and compares with."},{"cited_title":"Latency comparison of cloud datacenters and edge servers","cited_arxiv_id":null,"evidence_quote":"Charyyev et al. supplies the real-world cloud-to-edge latency ratios used to scale cloud latency in the evaluation."},{"cited_title":"Proofs of ownership in remote storage systems","cited_arxiv_id":null,"evidence_quote":"Halevi et al. provides the foundational proof-of-ownership protocol that PM-Dedup's challenge-response PoW builds on."},{"cited_title":"Boosting efficiency and security in proof of ownership for dedu- plication","cited_arxiv_id":null,"evidence_quote":"Di Pietro and Sorniotti supplies the efficiency-oriented PoW construction that motivates PM-Dedup's lightweight dual-level PoW."},{"cited_title":"An im- proved data stream summary: the count-min sketch and its applications","cited_arxiv_id":null,"evidence_quote":"Cormode and Muthukrishnan provides the Count-Min Sketch used for the lightweight share-index selection."},{"cited_title":"Message-locked encryption and secure deduplica- tion","cited_arxiv_id":null,"evidence_quote":"Bellare et al. defines message-locked encryption, the basis for deduplicating encrypted data with content-derived keys."}],"review_version":1}