Joongho Chang

Dr. Joongho Chang

orcid

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
3 July 2026 in Systems
Verified
Journal Article
Open access
PDF file
MKMinseong Kim
Minseong Kim
JCJoongho Chang
Joongho Chang
As artificial intelligence becomes increasingly embedded in healthcare operations, governance can no longer be treated solely as a static set of principles or regulatory requirements. This study proposes a Human-Centered Experience Engineering (HCEE)-based healthcare AI governance architecture designed as a closed-loop operational control system for stabilizing and adaptively managing AI-enabled healthcare systems. Rather than treating patient and employee experience as downstream outcomes, the framework repositions them as governance input signals, operationalized as the Patient Experience Index (PXI) and Employee Experience Index (EXI). The architecture integrates Policy, Governance, Control, and Experience through recurrent feedback, trigger-control rules, and adaptive learning mechanisms that translate experiential signals into ongoing operational adjustment. To examine how this architecture affects system behavior, agent-based simulation compares four scenarios: S1 (no governance), S2 (sense-only), S3 (full HCEE closed loop), and S4 (full HCEE under exogenous stress). Results show a consistent pattern of S1 < S2 < S3 in both PXI and EXI, indicating that sensing alone yields limited improvement, whereas feedback-coupled sensing, control, and learning produce stronger stabilization and performance gains. In S4, the same architecture demonstrates recovery, re-stabilization, and adaptive reinforcement under shock. These findings suggest that effective healthcare AI governance depends on whether feedback loops are architected to function operationally. These results are based on synthetic experience indices and are interpreted as conceptual, mechanism-level evidence.

27 August 2026 in Algorithms
Verified
Journal Article
Open access
PDF file
BJBONGJIN JUNG
BONGJIN JUNG
JCJoongho Chang
Joongho Chang
Voice phishing (vishing) causes financial and psychological harm, yet deployed detectors that scan transcripts for scam vocabulary are more fragile than their benchmark scores suggest: their near-perfect accuracy on standard Korean corpora is memorization, collapsing under scammer paraphrase or language-model rewriting. We construct KorMMP, a benchmark keeping real regulator-sourced scam audio and real benign speech but decorrelating transcript from label, and MiLTL, an on-device detector built on affective and neutrosophic channels, including a cross-modal contradiction signal (XM) formulated to weigh lexical warmth against vocal coldness. In the reported evaluation, XM’s measured contribution is confidence banding and escalation routing rather than ranking accuracy. A scoring rule with zero gradient-learned parameters at inference screens every segment, referring ambiguous calls to a small on-device language model; raw audio stays on the device. We measure both stages on a commodity CPU container and smartphone; these budgets cover the detector only and exclude speech recognition, which we expect to dominate a live end-to-end budget. With no training on the hard benchmark, MiLTL achieves an AUROC of 0.965 (recording-level cluster bootstrap 95% CI 0.943–0.984), while every evaluated comparator stays below 0.69: text-only, audio-only, fusion, 7B multimodal. MiLTL scores lower on the saturated corpus, as expected for a detector designed not to rely primarily on lexical shortcuts. KorMMP is a controlled stress test of lexical decorrelation, not a measure of in-the-wild detection; its harmful and benign audio come from different corpora, so the source and label are structurally confounded, and the residual source effects cannot be fully excluded. XM is author-defined, constructed rather than naturally occurring in the synthetic stratum and not yet validated against human perception; this is the principal open limitation of the work. Within that scope, the work provides a reproducible basis for on-device vishing defense and a benchmark for detector robustness under vocabulary shift.

30 July 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BJBONGJIN JUNG
BONGJIN JUNG
JCJoongho Chang
Joongho Chang
Voice phishing (vishing) causes massive financial and psychological harm worldwide, yet the detectors deployed against it are more fragile than their benchmark scores suggest. Systems that transcribe a call and scan it for scam vocabulary look near-perfect on standard Korean corpora, but that success is memorized: when a scammer rephrases the script—or rewrites it with a language model—accuracy collapses, while cloud transcription erodes privacy. We argue that manipulation should be caught by how a call is delivered, not which words it uses. We first build KorMMP, a benchmark that breaks this vocabulary shortcut: it keeps real regulator-sourced scam audio and real benign speech, then decorrelates the transcript from the label, so a keyword matcher cannot win. On it we present MiLTL, an on-device detector that reads a call in short segments through affective and neutrosophic voice channels—centered on a cross-modal contradiction signal that fires when warm words are spoken in a cold, controlling voice, a mismatch that text- and audio-only systems cannot see. A lightweight deterministic stage—an integer-weighted rule carrying zero learned parameters at inference—screens every segment and escalates only the calls that remain ambiguous to a small on-device language model, keeping the algorithm within a resource-constrained edge budget. Trained with the hard benchmark held out, MiLTL reaches AUROC 0.965 where text-only, audio-only, and fusion detectors stall below 0.69—the audio-only and fusion controls near chance—and—tellingly— scores lower on the saturated corpus, exactly as a detector that reasons about delivery rather than vocabulary memorization should. The result is a basis for on-device vishing defense that keeps raw audio local and stays robust when scam vocabulary shifts, together with a controlled benchmark for studying that robustness.

28 September 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BJBONGJIN JUNG
BONGJIN JUNG
JCJoongho Chang
Joongho Chang
Small language models can run locally when resource, cost or data-protection constraints rule out a larger model that could check their answers. We consider a quantized model whose runtime exposes tokens and log-probabilities but not hidden states. We propose MOME (Module of Metamorphic Evidence), a runtime gate that estimates the risk of a wrong answer without ground-truth labels at inference time. MOME edits each problem twice, once keeping the correct answer and once changing a condition so that the correct answer should change. A predictor fitted offline on labeled outputs of other backbones scores the resulting answers. A two-stage policy serves high-scoring answers, decides the rest by a three-vote majority or declines them. We tested MOME on mathematical word problems, whose definite answers make the correctness of every answer exactly measurable. On five quantized backbones and three benchmarks, the gate raised selective accuracy from to 69.5% to 85.8%. Correct yield was against (95% interval of the difference −0.6–+0.8 p) and met the preregistered point-estimate criterion for preservation. At the same coverage, the MOME score alone selected answers with accuracy, against for mean token log-probability. Adding the MOME score improved a linear probe on the hidden states of two backbones. Fitting held out each evaluated backbone but not its problems or benchmarks, and the results do not transfer directly to other tasks. Other domains can reuse the procedure: edits with known effects on the answer, a score fitted on other backbones and correct yield as the criterion.

31 October 2026 in Sensors
Verified
Journal Article
Open access
PDF file
BJBONGJIN JUNG
BONGJIN JUNG
JCJoongho Chang
Joongho Chang
Small language models often run locally, where no larger model can check their answers, and the runtime exposes tokens and log-probabilities but not hidden states. MOME (Module of Metamorphic Evidence) is a label-free runtime gate. It edits each problem twice by deterministic rules, once keeping the correct answer and once changing a condition on which that answer depends. A predictor fitted offline on other backbones scores the answers. A two-stage policy serves high-scoring answers and verifies the rest by a vote over further samples or declines them. Mathematical word problems were used because every answer can be checked exactly. On five four-bit backbones and three benchmarks, the gate raised selective accuracy from 69.5% to 85.8% with two stage-2 samples and to 88.1% with four, at 4.34 and 5.70 generations per item. Correct yield stayed at the ungated level. The MOME score ranked answers with an area under the ROC curve (AUROC) of 0.863, against 0.734 for mean token log-probability. On four backbones of further families, it reached 0.829 against 0.629, and correct yield changed by −1.81 percentage points at stage-2 temperature 0.7 and by −0.19 at 0.3. The fitted components apply to mathematical problems, but the procedure can be reused in other domains.