Source-linked AI summary
A Cognitive-Aware QML-CRL Framework for Detecting Affinity and Romance-Investment Fraud
Bibhas Adhikari, Ramya Srinivasan
TL;DR
Affinity and romance-investment fraud exploit trust and staged manipulation, raising the question of whether a system can detect the manipulation and determine when to act. The paper combines a trainable quantum-cognitive encoder with a classical reinforcement-learning stopping agent; on synthetic conversations, end-to-end training performs best under mis-specified initialization.
Problem
The paper asks whether a system can detect manipulation as it unfolds in affinity and romance-investment conversations and determine when to flag.
Source: §1 · pp. 1–2Method
A structured quantum circuit encodes cognitive-bias cues, framing order, and narrative coherence, while a classical reinforcement-learning agent learns whether to wait or flag.
Source: §1 · p. 2Results
Under mis-specified initialization, end-to-end training achieved F1 = 0.971 ± 0.041 with no false positives, while the frozen encoder missed all 62 fraud conversations.
Source: §4.1 · p. 10Takeaways & Limitations
The conclusion reports that the frame qubit proves necessary for detection in the evaluated synthetic conversations.
Source: §6 · p. 12The evaluation relies entirely on synthetic conversations built around the model’s funnel structure, and the NLP cue-mapping front end has not been independently validated.
Source: §5 · p. 12
Abstract
from arXiv · showhide
We present a hybrid quantum-classical framework that detects affinity and romance-investment fraud by modelling the cognitive biases in a manipulative conversation. In our proposed framework, cognitive biases central to this fraud class are carried by dedicated qubits in a structured parameterized quantum circuit, together with a frame qubit makes the encoding sensitive to the temporal order of manipulative reframing, and a narrative qubit that aggregates co-occurrence through a trainable entanglement layer. The circuit parameters are trained jointly with a classical reinforcement-learning agent that decides, turn by turn, whether to flag the conversation, modeled as an optimal stopping problem. We evaluate the model's performance on synthetic conversations that include hard negatives, legitimate but urgent, and legitimate but pushy sales conversations.
1 Introduction
Affinity and romance-investment fraud exploit trust and staged narratives, with cognitive biases operating in sequence rather than through a single deceptive claim. The paper proposes a learned quantum-cognitive encoder and classical stopping agent to detect manipulation and decide when to flag.
1 Introduction
- Affinity and romance-investment fraud exploit in-group trust or emotional intimacy through staged relationship-building and escalating financial requests.The paper describes these schemes as exploiting cognitive biases in a specific sequence rather than relying on one deceptive claim.Source: §1 · pp. 1–2
- Prior behavioral reinforcement learning embeds biases such as loss aversion, but its scalar utility transform does not reproduce order effects or the conjunction fallacy.Quantum-cognitive models instead represent beliefs as vectors and questions as generally non-commuting measurements.Source: §1 · p. 2
- The proposed QML-CRL model pairs a trainable quantum-cognitive encoder with a classical reinforcement-learning agent that learns when to flag.The encoder uses bias qubits, an order-carrying frame qubit, and a narrative qubit that aggregates co-occurrence.Source: §1 · p. 2
- Non-commuting frame-qubit rotations make the encoding depend on whether loss or recovery framing occurs first.The measurement basis must also avoid a degenerate choice that makes order dependence invisible; the narrative qubit aggregates coherence without exceeding the Fréchet bound.Source: §1 · pp. 2–3
- The classical agent treats detection as optimal stopping, trading off false alarms against intervening too late rather than classifying only after the conversation.The introduction says the numerical tests examine order-effect magnitude and isolate encoder-channel contributions to detection quality.Source: §1 · pp. 3–4
2 Quantum-Cognitive PQC Encoder
The encoder uses an (n + 2)-qubit register: bias qubits track cognitive cues, while frame and narrative qubits represent order and coherence. Turn-wise cue-conditioned operations produce risk features for the downstream detector.
2 Quantum-Cognitive PQC Encoder
- The (n + 2)-qubit register assigns n qubits to cognitive biases and uses frame and narrative qubits to monitor conversation order and coherence.The encoder is presented as the quantum component of an end-to-end architecture for the four biases described in the introduction.Source: §2 · p. 5
2.1 Quantum Register
- The per-turn cue vector drives the encoder, whose risk features are read from a readout-transformed copy of the evolving register state.The feature vector also includes a saturating time feature and an intercept independent of Θ.Source: §2.1 · p. 5
- Each bias qubit accumulates cue evidence through a soft OR, a differentiable relaxation of Boolean OR that saturates as the learned probability that a cue appeared on any prior turn.Each channel has trainable gain and threshold parameters rather than shared, hand-set values.Source: §2.1 · p. 6
- The frame qubit uses distinct Z- and Y-axis rotations for loss and recovery framing, making their temporal order observable except at a degenerate readout basis.At μ = ±π/2, the frame channel is provably order-blind.Source: §2.1 · p. 6
- Trainable controlled-RY gates connect each bias and the frame qubit to the narrative qubit, learning each channel’s contribution to narrative coherence.There are n + 1 controls, and setting φ_j = π makes each controlled rotation a CNOT.Source: §2.1 · p. 6
- For n cognitive biases, the circuit producing st uses n + 2 qubits and comprises t(n + 2) + 1 rotations and n + 1 two-qubit controlled rotations.Source: §2.1 · p. 6
3 Joint Quantum-Classical Learning
The learning framework couples the quantum encoder to a classical agent that observes its risk features and chooses whether to wait or flag. The agent learns through rewards for timely detection and penalties for false alarms or missed fraud.
3 Joint Quantum-Classical Learning
- The RL agent observes the encoder’s risk features and chooses WAIT, which advances the conversation, or FLAG, which terminates the episode.The paper describes this decision process as part of a hybrid quantum-classical learning framework.Source: §3 · pp. 6–7
- The reward design penalizes false alarms and missed fraud while discounting late-but-correct detections, with a 20/80 fraud/benign base rate.At a 50/50 rate with P_fa = 5, blind flagging would yield expected reward +2.5 and become the reward-function optimum.Source: §3 · p. 7
- The action-value function is linear in the encoder output, making the model a PQC feature map followed by a two-row linear head.The paper places representational capacity beyond that head in the circuit.Source: §3 · p. 7
- The training procedure updates the classical head and circuit parameters using the semi-gradient temporal-difference error with the bootstrap target held constant.The full procedure is given in Algorithms 1 and 2, while gradient-learning details are deferred to the supplementary section.Source: §3 · p. 8
4 Numerical Simulations with Synthetic Data
The simulations test the framework on staged synthetic conversations with hard negatives, then assess joint training and the contribution of the frame channel. Under mis-specified initialization, end-to-end training outperforms a frozen encoder, while frame information distinguishes fraud from pushy sales.
4 Numerical Simulations with Synthetic Data
- Synthetic data uses six staged fraud phases, slower, lower-intensity SUBTLEFRAUD, and BENIGN, LEGITURGENT, and bias-overlapping PUSHYSALES negatives.The test and training sets each contain 300 conversations, with a 20% fraud base rate; results are averaged over three seeds.Source: §4 · pp. 8–10
4.1 Simulation results
- Removing the frame qubit collapses F1 to 0.000; after training, its feature separates fraud from PUSHYSALES while loss and authority saturate across all three classes.The frame feature rises from −1 to −0.08 after the loss event for fraud, while remaining at −1 for both negative classes.Source: §4.1 · pp. 10–11
- F1 = 0.971 ± 0.041 with no false positives under mis-specified initialization, versus F1 = 0.000 for the frozen encoder, which missed all 62 fraud conversations.With cognitive initialization, trained and frozen encoders achieved F1 = 0.934 ± 0.019 and F1 = 0.919 ± 0.051, respectively; the difference was not considered decisive.Source: §4.1 · p. 10, Table 2 · p. 11
- Figure 3 evaluates detection timing after cue-preserving turn-order shuffling, and reports that joint training enables post-warm-up adaptation while the frozen encoder fails to flag.Detection timing is sequence-dependent at the decision level.Source: Figure 3 · p. 11, §4.1 · p. 11
5 Limitations and Future Work
The framework has hardware-cost, training-stability, and evaluation-scope limitations. The evaluation uses synthetic conversations, and the NLP front end for converting text into cues has not been independently implemented or validated.
5 Limitations and Future Work
- A T-turn hardware trajectory requires T independent circuit executions, costing O(T × shots) per conversation; training adds shifted evaluations and shot noise.The authors leave systematic shot-noise analysis for future work.Source: §5 · pp. 11–12
- Joint optimization can be fragile because semi-gradient TD learning updates a moving feature map; warm-up, anchoring, and adaptive gradient scaling are required for stability.Anchor strength trades stability against adaptability; policy-gradient methods or fitted-Q learning with a target network are suggested alternatives.Source: §5 · p. 12
- Evaluation relies entirely on synthetic conversations, and the NLP front end is specified but not independently implemented or validated; cue-extraction errors would propagate to the circuit.Source: §5 · p. 12
6 Conclusion
The proposed hybrid pipeline combines a structured quantum circuit for manipulation-relevant biases and narrative cues with a classical reinforcement-learning agent. On synthetic conversations with hard negatives, end-to-end training makes the encoder the best-performing configuration, and the frame qubit proves necessary for detection.
6 Conclusion
- On synthetic conversations with hard negatives, end-to-end training makes the encoder the best-performing configuration, while a frozen or mis-specified encoder is useless.The frame qubit also proves necessary for detection.Source: §6 · p. 12
- The pipeline jointly trains a structured parameterized quantum circuit and a classical reinforcement-learning agent that learns when to flag conversations.The circuit encodes manipulation-relevant biases, an order-sensitive frame qubit, and a narrative-coherence aggregator.Source: §6 · p. 12
- The frame qubit proves necessary for detecting manipulation in the evaluated synthetic conversations.Source: §6 · p. 12
7 Supplementary
Supplementary material formalizes the frame and narrative readouts, specifies gradient and stopping-agent updates, describes stabilization measures, and details synthetic conversation generation.
7.1 Mathematical Foundations
- Mixed RY/RZ gates make the frame state order-dependent, but readout is blind at µ = ±π/2; the authors train µ and initialize it at −π/4.Source: §7.1 · pp. 16–17
- The narrative qubit aggregates coherence without violating the joint-measurement Fréchet bound, while frame-qubit order sensitivity is carried separately; the feature trajectory is classically evaluable in O(nT).Quantum-hardware implementation is left as future work.Source: §7.1 · pp. 16–17
- At φ = π, narrative contrast is a local maximum with zero first-order signal, so entangler angles cannot move from that initialization; initializing away avoids the degeneracy.Source: §7.1 · p. 18
7.2 Gradients through the circuit and the joint training
- Eq. (10) is the sole interface: agent surprise δt projects through its weights onto Born-rule Jacobians; simulations compute derivatives analytically and validate them to 1.0 × 10−9.Source: §7.2 · pp. 18–19
7.3 Algorithm 2
- The stopping agent receives only register readouts, while the circuit receives only δt; joint training therefore makes the encoder support stopping decisions rather than classify turns.The objective is optimal stopping, with mean episode return reported rather than training loss.Source: §7.3 · pp. 21–22
7.3.1 Stability of the joint update mechanism
- Joint TD updates sometimes collapse to never-flag or always-flag policies; warm-up, anchoring, and episode-averaged RMSProp-scaled steps are used to stabilize learning.Source: §7.3.1 · p. 22
7.4 Synthetic data generation
- Fraud traces span six stages, with one loss-to-recovery frame transition confined to the last two and independent per-turn noise (σ = 0.10), not autoregressive structure.For T = 14, the critical turn is the first turn of the loss-event stage, turn 9.Source: Figure 5 · p. 23, §7.4 · p. 23