Ashutosh Agarwal

Mr. Ashutosh Agarwal


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

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

Representative Publications

View All Publications
27 September 2025 in International Journal of Applied Mathematics
Verified
Journal Article
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
NeuroFusion-X is a unified, modular framework for end-to-end prediction from heterogeneous real-world data. Many decisions require joint reasoning over time-series, images, and text, yet production systems remain siloed, and naïve early/late fusion misses cross-modal dependencies and temporal alignment. NeuroFusion-X addresses this via: (1) modality-specialized encoders,CNNs for images, a compact transformer for text, and a bidirectional time-series encoder with temporal attention; (2) a cross-modal fusion-attention block that learns instance-wise interactions and down-weights noisy or missing channels; and (3) parameter-efficient bottlenecks and inference-oriented kernels to cut latency without sacrificing accuracy. To evaluate realism and scale while avoiding privacy constraints, we construct a controlled synthetic benchmark of 500k multimodal samples across healthcare, finance, and cybersecurity. Each sample includes a 48-step, 30-variable time-series, a 128×128 image, and a 60–160-token note, with class imbalance, inference-time modality masks, and induced distribution shifts. Across 18 tasks, NeuroFusion-X reaches approximately 97.8% mean accuracy and approximately 0.976 macro-F1, reducing median per-sample inference latency by approximately 35% versus a strong baseline. Robustness holds with ≤1.6% macro-F1 drop under 20% modality dropout and ≤2.2% under light adversarial perturbations. Ablations show fusion-attention, modality-dropout, and domain-adaptive normalization drive reliability. We outline deployment pathways for safety-critical contexts and integration with multimodal LLMs for rationale-grounded predictions.

28 November 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
This paper introduces a novel framework for generating counterfactual explanations that prioritize faithfulness to the underlying model behavior. The approach leverages energy-based modeling to constrain the generation of counterfactuals, ensuring they align with the model’s learned data distribution. By integrating this energy constraint with conformal prediction, the methodology achieves a balance between faithfulness and plausibility. Empirical evaluations demonstrate the effectiveness of the proposed framework in producing counterfactuals that accurately reflect model decision boundaries, offering improved interpretability for complex machine learning models. The method provides a way to distinguish trustworthy from unreliable models by focusing on the energy constraint.

27 November 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
This study presents an analytical evaluation of autonomous computational agents designed to solve complex, logicintensive challenges in dynamic digital environments. The framework explores how reasoning-based algorithms, adaptive learning mechanisms, and multi-step decision strategies can be employed to achieve goal-oriented automation without domainspecific pretraining. Through systematic experimentation, the research examines agentic behavior under variable task conditions, emphasizing adaptability, robustness, and problem decomposition efficiency. Empirical analysis demonstrates that structured reasoning and iterative self-correction can enhance performance consistency across diverse computational tasks. The findings contribute to advancing intelligent agent architectures in computer science, offering insights into scalable, general-purpose systems for autonomous decision-making and digital security problem solving.

10 December 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
This paper proposes and evaluates a unified machine-learning framework for enterprise portfolio management that integrates multi-horizon financial forecasting, unsupervised risk detection, and explainable reporting within a single pipeline. Using a synthetic but structurally realistic ERP-style dataset comprising 162,000 project–month records with 24 financial and operational features, the study adopts a quantitative design based on multi-source feature engineering, expanding-window temporal cross-validation, and benchmarking of five forecasting models (Linear Regression, Random Forest, XGBoost, LightGBM, CatBoost) across 1-, 3-, and 6-month horizons. Hyperparameters for the strongest models are tuned with Optuna, and three unsupervised detectors (Isolation Forest, COPOD, LODA) are applied to scaled numeric features, while SHAP is used to generate global and local explanations. Results show that gradient-boosted trees substantially outperform linear baselines, reducing MAE by roughly 25–40% and achieving R² ≈ 0.63 at 1 month, ≈ 0.57 at 3 months, and ≈ 0.43 at 6 months, with open commitments, backlog, change orders, and schedule slippage emerging as dominant drivers of future spend. The anomaly layer flags around 2% of records as high risk, capturing patterns such as vendor rate spikes, zero-commitment overspend, stalled backlog, and abrupt forecast collapses. Rather than introducing novel algorithms, the contribution of this work lies in a unified, SHAP-enabled architecture that enhances auditability and governance by turning model outputs into defensible financial narratives and providing a practical blueprint that future work can extend to real ERP data, streaming architectures, and human-in-the-loop risk governance.

21 November 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
The possibility and effect of cyberattacks have increased due to the expansion of enterprise attack surfaces brought about by the acceleration of digital transformation. The dynamic, high-dimensional signals present in contemporary threat intelligence are difficult for traditional actuarial and rule-based risk-rating techniques to capture. We present TriRisk, a multi-task deep neural network that simultaneously forecasts (i) the likelihood of a substantial breach within the next 12 months, (ii) the anticipated financial severity, and (iii) the anticipated time-to-breach. trained using our recently assembled CyberThreat-Enterprise-2025 dataset (five years, 4,886 firms, 118,204 incidents, 312 characteristics), TriRisk achieves AUROC 0.892, normalized MAE 0.208 for severity, and MAPE 15.2% for time-to-breach on a 2024 out-of-time test set, outperforming state-of-the-art baselines. While robustness tests against FGSM and PGD adversarial perturbations show less than 3% performance deterioration, SHAPbased explainability indicates susceptibility to remote-service vulnerabilities and supplier compromise as key causes. By improving quantitative cyber-risk assessment, the framework influences regulatory capital modeling, cyber-insurance pricing, and security investment prioritization.

12 December 2025 in Preprints.org
Verified
Preprint
Open access
PDF file
AAAshutosh Agarwal
Ashutosh Agarwal
Large Language Models (LLMs) are increasingly deployed in Security Operations Centers (SOCs) for alert triage and threat‑intelligence synthesis. We study Adversarial Hallucination Engineering (AHE): attacks that bias LLM reasoning by introducing small clusters of poisoned context into retrieval‑augmented generation (RAG) pipelines, producing targeted fabrications aligned with attacker goals. Using a safe, fully synthetic simulator of a RAG+LLM SOC, we formalize the AHE threat model, introduce Hallucination Propagation Chains (HPCs)—mutually reinforcing poisoned documents designed to create artificial consensus at retrieval time—and evaluate a lightweight defense, Chain‑of‑Thought Attestation (CoTA), based on per‑token uncertainty, provenance attribution, and source reputation. Across three model scales, hallucination‑induction rate (HIR) rises superlinearly with HPC size (e.g., for a Large‑70B proxy, 12.45%→61.84% as HPC size in top‑k grows 0→5); actionable misconfiguration rate (AMR) grows from 3.23% (no attack) to 38.18% (HPC‑5). CoTA reduces attack‑success rate (ASR) by ~55% for HPCs≥3 at ~7% false‑positive flags and ~8% latency overhead. We release synthetic artifacts to support reproducible, defensive research.