The Accuracy–Fairness–Efficiency Trilemma in Mobile Image Classification: A Pareto Benchmark

The Accuracy–Fairness–Efficiency Trilemma in Mobile Image Classification: A Pareto Benchmark | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article The Accuracy–Fairness–Efficiency Trilemma in Mobile Image Classification: A Pareto Benchmark Van Thanh Tran, Thanh Hien Lam, Nang Toan Do, Tuan-Tu Huynh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9321945/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 14 You are reading this latest preprint version Abstract Deploying deep learning classifiers on resource-constrained mobile devices requires the simultaneous satisfaction of three structurally competing objectives: predictive accuracy, demographic fairness, and inference efficiency. Although each objective has been studied independently, no prior work has (i) formalized their joint structure as a constrained multi-objective optimization problem with normalized, dimensionally consistent component losses; (ii) provided a systematic Pareto benchmark over a comprehensive set of optimization strategies under identical experimental conditions; or (iii) defined a Deployment-Feasible Zone (DFZ) as the feasibility-constrained subset of the Pareto frontier. This paper provides all three contributions. We formalize the Accuracy–Fairness–Efficiency Trilemma with hard deployment constraints (F1 ≥ 0.85, EOD < 0.10 per protected attribute, model size ≤ 10 MB, latency ≤ 300 ms on an entry-level SoC) and benchmark eleven optimization configurations spanning 2D augmentation, 3D-aware augmentation (DECA parametric space), fairness-constrained pruning (PFP), INT8 quantization, knowledge distillation to MobileNetV3, and their combinations. Experiments use a 2,821-image facial classification dataset with 24 intersectional demographic subgroups (worst-case imbalance 35.47:1), selected as a demanding stress-test; the formalization is designed to be portable to other mobile classification tasks satisfying analogous structural conditions (Section 6.5). Seven of eleven configurations qualify for the DFZ; the combination of 3D-aware augmentation and Protected Fairness Pruning (C2) is the Pareto knee point: F1 = 0.934 (95 % CI: 0.906–0.962), EODgender = 2.0 %, 6.3 MB, 187 ms. Standard magnitude pruning is formally shown to be Pareto-dominated by fairness-constrained pruning at identical compression ratio. The Adaptive Trilemma Weight Scheduler (ATWS) yields consistent gains of 1.3 pp F1 and 0.6–0.7 pp EOD reduction; bootstrap stability analysis (n = 1,000) confirms C2 as the knee point in 94.2 % of resamples. Physical sciences/Engineering Physical sciences/Mathematics and computing multi-objective optimization Pareto frontier demographic fairness model compression knowledge distillation quantization 3D data augmentation image classification edge deployment Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 04 May, 2026 Reviews received at journal 02 May, 2026 Reviews received at journal 28 Apr, 2026 Reviews received at journal 25 Apr, 2026 Reviewers agreed at journal 25 Apr, 2026 Reviewers agreed at journal 25 Apr, 2026 Reviews received at journal 21 Apr, 2026 Reviewers agreed at journal 21 Apr, 2026 Reviewers agreed at journal 20 Apr, 2026 Reviewers invited by journal 20 Apr, 2026 Editor assigned by journal 19 Apr, 2026 Editor invited by journal 17 Apr, 2026 Submission checks completed at journal 10 Apr, 2026 First submitted to journal 10 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9321945","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":629696012,"identity":"c65baba3-51f1-4aec-9ffa-4ecf236f2fb4","order_by":0,"name":"Van Thanh Tran","email":"","orcid":"","institution":"Lac Hong University","correspondingAuthor":false,"prefix":"","firstName":"Van","middleName":"Thanh","lastName":"Tran","suffix":""},{"id":629696013,"identity":"0c7dee0b-ebe4-476c-b6d8-5bd86b90c16a","order_by":1,"name":"Thanh Hien Lam","email":"","orcid":"","institution":"Lac Hong University","correspondingAuthor":false,"prefix":"","firstName":"Thanh","middleName":"Hien","lastName":"Lam","suffix":""},{"id":629696014,"identity":"ed5db9c0-98bc-4630-9c6f-35fdd843d456","order_by":2,"name":"Nang Toan Do","email":"","orcid":"","institution":"Vietnam Academy of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Nang","middleName":"Toan","lastName":"Do","suffix":""},{"id":629696015,"identity":"196ec832-7970-47d1-944d-61d917971116","order_by":3,"name":"Tuan-Tu Huynh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA30lEQVRIiWNgGAWjYBACAyBmhrIZHzxsANEJxGgBKWJjYDZIJFULmwRRWswlcsweF/44LG8u32NWkbjjMAM/e44BM+8O3FosZ+SYG89IOGy4s43H7EbimcMMkj1vgFrO4HHYjRwzaZ6Ew4wbjoG0tB0GiRgwzmwjrMUepKUApMWeWC2JIC0MYFskcgwYPuLTcuZZmTRPWnryhmNpxRKJbek8EmeeFRzAq+V48jZpHhtr2w2HD2/88LHNWo6/PXnjg0Q8WhgYOAxQuDwg4gA+DQwM7A/wy4+CUTAKRsEoAAAc+VEFQzz6rgAAAABJRU5ErkJggg==","orcid":"","institution":"Lac Hong University","correspondingAuthor":true,"prefix":"","firstName":"Tuan-Tu","middleName":"","lastName":"Huynh","suffix":""}],"badges":[],"createdAt":"2026-04-04 16:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9321945/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9321945/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108490914,"identity":"9a2e64c5-3b66-4d8b-8985-f673f5c12ed5","added_by":"auto","created_at":"2026-05-05 09:49:51","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":811113,"visible":true,"origin":"","legend":"","description":"","filename":"trilemmaRevised.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9321945/v1_covered_a2db2516-5694-4439-af0c-a45e4479ebc0.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"The Accuracy–Fairness–Efficiency Trilemma in Mobile Image Classification: A Pareto Benchmark","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"multi-objective optimization, Pareto frontier, demographic fairness, model compression, knowledge distillation, quantization, 3D data augmentation, image classification, edge deployment","lastPublishedDoi":"10.21203/rs.3.rs-9321945/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9321945/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDeploying deep learning classifiers on resource-constrained mobile devices requires the simultaneous satisfaction of three structurally competing objectives: predictive accuracy, demographic fairness, and inference efficiency. Although each objective has been studied independently, no prior work has (i) formalized their joint structure as a constrained multi-objective optimization problem with normalized, dimensionally consistent component losses; (ii) provided a systematic Pareto benchmark over a comprehensive set of optimization strategies under identical experimental conditions; or (iii) defined a Deployment-Feasible Zone (DFZ) as the feasibility-constrained subset of the Pareto frontier. This paper provides all three contributions. We formalize the Accuracy–Fairness–Efficiency Trilemma with hard deployment constraints (F1 ≥ 0.85, EOD \u0026lt; 0.10 per protected attribute, model size ≤ 10 MB, latency ≤ 300 ms on an entry-level SoC) and benchmark eleven optimization configurations spanning 2D augmentation, 3D-aware augmentation (DECA parametric space), fairness-constrained pruning (PFP), INT8 quantization, knowledge distillation to MobileNetV3, and their combinations. Experiments use a 2,821-image facial classification dataset with 24 intersectional demographic subgroups (worst-case imbalance 35.47:1), selected as a demanding stress-test; the formalization is designed to be portable to other mobile classification tasks satisfying analogous structural conditions (Section 6.5). Seven of eleven configurations qualify for the DFZ; the combination of 3D-aware augmentation and Protected Fairness Pruning (C2) is the Pareto knee point: F1 = 0.934 (95 % CI: 0.906–0.962), EODgender = 2.0 %, 6.3 MB, 187 ms. Standard magnitude pruning is formally shown to be Pareto-dominated by fairness-constrained pruning at identical compression ratio. The Adaptive Trilemma Weight Scheduler (ATWS) yields consistent gains of 1.3 pp F1 and 0.6–0.7 pp EOD reduction; bootstrap stability analysis (n = 1,000) confirms C2 as the knee point in 94.2 % of resamples.\u003c/p\u003e","manuscriptTitle":"The Accuracy–Fairness–Efficiency Trilemma in Mobile Image Classification: A Pareto Benchmark","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-28 15:37:42","doi":"10.21203/rs.3.rs-9321945/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-04T05:37:23+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-02T17:09:47+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-28T15:30:55+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-25T18:11:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"48140283272358496153196216544573687701","date":"2026-04-25T17:52:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"156218730900754700500511745510071278787","date":"2026-04-25T13:56:56+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-21T06:50:32+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"92153321827761790295593139552411021122","date":"2026-04-21T06:25:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"270092047258829037863521348828102696009","date":"2026-04-21T02:50:21+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-20T12:46:14+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-19T15:03:25+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-17T14:59:02+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-10T09:48:47+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-04-10T09:03:07+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"99063675-8279-4154-b9d9-e05b36aec344","owner":[],"postedDate":"April 28th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-04T05:37:23+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-02T17:09:47+00:00","index":47,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67023382,"name":"Physical sciences/Engineering"},{"id":67023383,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-05-14T08:08:18+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-28 15:37:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9321945","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9321945","identity":"rs-9321945","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}