---
title: Collision-Alleviated Signal Hash (CASH)
url: https://www.emergentmind.com/topics/collision-alleviated-signal-hash-cash
type: topic
---

# Collision-Alleviated Signal Hash (CASH)

Searching arXiv for the CASH signal-hashing paper and related acronym usages to ground the article in current arXiv literature.
Collision-Alleviated Signal Hash (CASH) is a hash-based model for Online Specific Emitter Identification (OSEI), a setting in which a model is constructed from signal samples from seen emitters and must then identify new samples from both seen and novel emitters online during inference [2509.23807]. In this formulation, each incoming complex-valued IQ signal is mapped to a compact binary hash code, augmented with a 1-bit seen-emitters indicator, and resolved to an emitter identity by hash-table lookup. The method is designed for scenarios in which novel emitters may appear after deployment, labeled samples from those emitters may be few or absent, and the system must operate without retraining while processing signals one by one.

## 1. OSEI problem setting and formal scope

Specific Emitter Identification (SEI) seeks to distinguish individual devices that use the same communication standard and protocol by exploiting device-unique hardware imperfections, often termed Radio Frequency Fingerprints (RFFs). In the formulation associated with CASH, conventional SEI assumptions are relaxed: test-time signals need not come only from the training emitters, new emitters may appear after deployment, and inference must be online rather than batch-based.

The OSEI setting distinguishes **seen emitters**, whose labeled signals are available during training, from **novel emitters**, which appear at test time. The notation in the underlying formulation uses \(Y^s\) for the training class space of seen emitters, \(Y^n\) for novel emitters, and \(\mathcal{Y}^{Te} = Y^s \cup Y^n\) for the test-time label space. In the generalized zero-shot variant, the training set is
$$
D^{Tr} = \{ (\mathbf{r}_i, y_i) \mid \mathbf{r}_i \in \mathcal{R}, y_i \in Y^s \}_{i=1}^{N},
$$
while the test set is
$$
D^{Te} = \{ (\mathbf{r}_j, y_j) \mid \mathbf{r}_j \in \mathcal{R}, y_j \in \mathcal{Y}^{Te} \}_{j=1}^{J}.
$$
The objective is to learn a function \(f: \mathcal{R} \to \mathcal{Y}^{Te}\) that outputs an emitter identity for each arriving signal.

Two sub-tasks are defined. **Online Few-Shot Learning (FSL)** uses abundant labeled samples from many seen emitters together with a small number of labeled samples from each novel emitter of interest; test-time signals come from novel emitters and must be identified online at per-emitter granularity. **Online Generalized Zero-Shot Learning (GZSL)** trains only on seen emitters and tests on a mixture of seen and novel emitters in task-agnostic mode, without an oracle revealing whether a test signal is from \(Y^s\) or \(Y^n\).

This formulation is motivated by two limitations of earlier SEI methods. Open-set methods can place out-of-distribution signals into a single “unknown” class but do not distinguish which novel emitter produced them. Zero-shot and few-shot methods that rely on offline clustering or fine-tuning are not naturally suited to streaming inference and tend to bias predictions toward the seen-emitter distribution.

## 2. CASH architecture and online identification pipeline

CASH decomposes online identification into a **seen emitters identifying step** and a **signal hash coding step**. The first step determines whether a signal likely originates from a seen or novel emitter; the second assigns a compact binary code intended to identify the specific emitter.

The **signal encoder** takes a sliced IQ signal and produces a latent embedding. The implementation described for CASH uses a **9-layer complex-valued CNN (CVCNN)** with an added dense layer of dimension 768 as the embedding extractor \(\mathcal{E}\). Given an augmented input \(\tilde{\mathbf{r}}\), the encoder outputs
$$
\mathbf{e} = \mathcal{E}(\tilde{\mathbf{r}}), \qquad \mathbf{e} \in \mathbb{R}^{768}.
$$
During training, an **embedding enhancer** \(\mathcal{A}\), implemented as a 3-layer MLP with output dimension 12, produces \(\mathbf{z}_p = \mathcal{A}(\mathbf{e}_p)\) for supervised contrastive learning.

The **online signal hasher** contains two projectors that share a 3-layer fully connected trunk. The **sign projector** \(\mathcal{H}_s\) produces attribute indicators, and the **confidence projector** \(\mathcal{H}_c\) produces per-bit confidence magnitudes. At inference time, the binary hash is
$$
\mathbf{h} = \operatorname{sign}(\mathcal{H}_s(\mathbf{e})) \in \{-1,1\}^{F},
$$
where \(F\) is the hash length. During training, non-differentiable operations are replaced with tanh-based approximations:
$$
\mathbf{h}_p = \tanh(\mathcal{H}_s(\mathbf{e}_p)),
$$
$$
\mathbf{c}_p = \tanh(\mathcal{H}_c(\mathbf{e}_p)) \otimes \mathcal{H}_c(\mathbf{e}_p),
$$
and the combined feature used for contrastive training is
$$
\mathbf{z}'_p = \mathbf{h}_p \otimes \mathbf{c}_p.
$$

The **seen emitters identifier** is based on **Adversarial Reciprocal Points Learning (ARPL)**. A projector \(\mathcal{O}\) maps \(\mathbf{e}\) to an \(n\)-dimensional space with \(n=128\):
$$
\mathbf{z} = \mathcal{O}(\mathbf{e}).
$$
For each seen class \(u \in Y^s\), a reciprocal point \(\mathbf{P}_u \in \mathbb{R}^n\) is defined. The distance used by the identifier is
$$
d(\mathbf{z}, \mathbf{P}_u)
= \frac{1}{n}\|\mathbf{z} - \mathbf{P}_u\|_2^2 - \mathbf{z}\cdot \mathbf{P}_u.
$$
At inference time, CASH computes
$$
m = \max_{u \in Y^s} d(\mathbf{z}, \mathbf{P}_u),
$$
compares \(m\) to a threshold \(T\), and outputs an indicator
$$
a =
\begin{cases}
1, & \text{if } m > T,\\
-1, & \text{if } m \le T.
\end{cases}
$$
This 1-bit output is concatenated with the binary hash to form the **collision-alleviated hash code**
$$
\mathbf{h}^{\mathrm{c}} = [\mathbf{h}, a].
$$

Online identity assignment is then reduced to lookup in a hash table \(S\). If \(\mathbf{h}^{\mathrm{c}}\) is absent from \(S\), the system inserts the code and assigns a new emitter identity; otherwise it reuses the stored identity:
$$
\hat{y} =
\begin{cases}
\text{Ind}(S, \mathbf{h}^{\mathrm{c}}), & \mathbf{h}^{\mathrm{c}} \in S,\\
\text{assign new ID, then } \text{Ind}(S, \mathbf{h}^{\mathrm{c}}), & \text{otherwise}.
\end{cases}
$$
Only the hash table is updated online; the neural modules remain fixed after offline training.

## 3. Objective functions and collision-alleviation mechanism

CASH trains the encoder, identifier, and hasher in two stages. The first stage jointly optimizes the encoder and seen-emitters identifier with
$$
\mathcal{L}_1 = \mathcal{L}_{\mathrm{E}} + \alpha \mathcal{L}_{\mathrm{I}},
$$
where \(\mathcal{L}_{\mathrm{E}}\) is a supervised contrastive loss over encoder embeddings and \(\mathcal{L}_{\mathrm{I}}\) is the ARPL-based identifier loss. The second stage trains the hasher on top of the fixed encoder using
$$
\mathcal{L}_2 = \mathcal{L}_{\mathrm{H}} + \beta \mathcal{L}_{\mathrm{R}},
$$
where \(\mathcal{L}_{\mathrm{H}}\) is a supervised contrastive loss on the attribute-confidence representation \(\mathbf{z}'_p\), and \(\mathcal{L}_{\mathrm{R}}\) combines two explicit regularizers.

The identifier loss has two terms. First, the ARPL classifier uses a softmax over distances:
$$
p(u \mid \mathbf{e}_p)
=
\frac{\exp(d(\mathbf{z}_p, \mathbf{P}_u))}
{\sum_{v \in Y^s}\exp(d(\mathbf{z}_p, \mathbf{P}_v))}.
$$
Second, an adversarial margin constraint
$$
\mathcal{L}_{\mathrm{AMC}}
=
\frac{1}{2N}\sum_{p=1}^{2N}\max\bigl(d_s(\mathbf{z}_p, \mathbf{P}_{y_p}) - R, 0\bigr)
$$
pushes seen-sample embeddings to the boundary of a bounded region. The total identifier loss is
$$
\mathcal{L}_{\mathrm{I}} = \mathcal{L}_{\mathrm{CE}} + \lambda \mathcal{L}_{\mathrm{AMC}}.
$$

Collision alleviation in CASH is implemented through four interacting mechanisms. First, the supervised contrastive objective on \(\mathbf{z}'_p\) makes same-emitter representations similar and different-emitter representations dissimilar. Second, the **binary constraint**
$$
\mathcal{C}_{\mathrm{BIN}}
=
\frac{1}{2N}\sum_{p=1}^{2N}\left(1 - \frac{\|\mathbf{h}_p\|_1}{F}\right)
$$
encourages hash activations to approach \(\pm 1\), reducing discretization instability around zero. Third, the **similarity constraint**
$$
\mathcal{C}_{\mathrm{SIM}}
=
-\sum_{p=1}^{2N}\frac{1}{|I(p)|}\sum_{q \in I(p)}
\log
\frac{\exp(\mathbf{h}_p \cdot \mathbf{h}_q)}
{\sum_{k \ne p}\exp(\mathbf{h}_p \cdot \mathbf{h}_k)}
$$
acts directly in hash space, enforcing high inner products for same-emitter codes and lower inner products for different-emitter codes. Fourth, the appended indicator bit \(a\) partitions code space into seen and novel regions; even if a seen emitter and a novel emitter share the same \(\mathbf{h}\), their final codes \(\mathbf{h}^{\mathrm{c}}\) differ in the last bit [2509.23807].

In this construction, collisions arise from limited hash length, intra-class variability induced by noise or channel effects, and training bias toward seen emitters. CASH addresses these sources by separating robust sign information from per-bit confidence, by regularizing codes toward stable binarization, and by explicitly separating seen and novel emitters at the code level.

## 4. Data regime, training protocol, and reported performance

CASH is evaluated on two real-world signal datasets. **ADSB-10** contains 10 aircraft ADS-B emitters with center frequency 1090 MHz, 50 MHz sampling rate, 2-PPM modulation, and IQ dimension \(2 \times 4800\). **ORACLE-16** contains 16 Wi-Fi USRP X310 transmitters with center frequency 2450 MHz, 5 MHz sampling rate, 802.11 OFDM, and IQ dimension \(2 \times 6000\). In both cases, center slicing or random slicing is used to produce segments of length \(l=4000\).

For GZSL, ADSB-10 is split into 5 seen and 5 novel emitters, and ORACLE-16 into 8 seen and 8 novel emitters. For FSL, the ADSB-10 setting uses 90 other ADSB emitters as seen classes and all 10 ADSB-10 emitters as novel classes, while the ORACLE-16 setting uses 10 seen emitters and the remaining 6 as novel classes. GZSL training uses Adam with learning rate \(10^{-3}\) for 500 epochs and then SGD with learning rate \(10^{-4}\) for 100 epochs, both with batch size 128. The reported hyperparameters are \(\alpha = 0.1\), \(\beta = 100\) for ADSB-10, and \(\alpha = 10^{-4}\), \(\beta = 15000\) for ORACLE-16; the seen/novel threshold uses confidence \(\gamma = 0.95\). Hash length is set to \(F=12\) for GZSL and \(F=5\) for FSL. In FSL, the encoder is first pretrained with SimCLR for 300 epochs, then trained with \(\alpha = 0\) and batch size 10 for 300 epochs [2509.23807].

The principal GZSL results are summarized below.

| Setting | Best baseline | CASH |
|---|---:|---:|
| ADSB-10, task-aware All | 80.25 | **84.21** |
| ADSB-10, task-agnostic All | 58.84 | **70.07** |
| ORACLE-16, task-aware All | 86.85 | **87.98** |
| ORACLE-16, task-agnostic All | 67.36 | **75.91** |

The same experiments report detailed seen/novel breakdowns. On ADSB-10 task-aware GZSL, CASH achieves **99.44 / 68.98** on seen/novel emitters, compared with **93.32 / 67.18** for the best baseline. On ADSB-10 task-agnostic GZSL, CASH achieves **94.04 / 46.10**, compared with **89.36 / 28.32**. On ORACLE-16 task-aware GZSL, CASH achieves **98.69 / 77.26**, compared with **96.79 / 76.91**. On ORACLE-16 task-agnostic GZSL, CASH achieves **92.42 / 59.40**, compared with **86.21 / 48.51** or **88.76 / 44.60** for the strongest baselines, yielding the reported **+8.55\%** improvement in overall task-agnostic accuracy.

For online few-shot SEI, CASH consistently outperforms SA2SEI by **\(\ge 6.08\%\)** on ADSB-10 and **\(\ge 6.46\%\)** on ORACLE-16 across training regimes from 5 to 50 samples per novel emitter. With only 5 samples per novel emitter, the ORACLE-16 result is reported as **\(\approx 90.94\%\)** accuracy.

Ablation experiments isolate the collision-alleviation components. Adding sign and confidence projectors raises task-aware “All” accuracy from **61.85 to 67.88** on ADSB-10 and from **74.79 to 77.29** on ORACLE-16. Adding regularization raises these to **82.80** and **86.89**. Adding the seen-emitters identifier raises task-agnostic “All” accuracy from **54.17 to 70.07** on ADSB-10 and from **64.41 to 75.91** on ORACLE-16. Confusion matrices further indicate that without the seen-emitters identifier, collision rates between seen and novel emitters are approximately **38.9\%** on ADSB and **31.62\%** on ORACLE, whereas with the indicator they are reduced to **14.90\%** and **12.94\%**.

## 5. Terminological position and relation to other hashing literatures

The acronym **CASH** is not unique to online SEI. In password security, **“CASH: A Cost Asymmetric Secure Hash Algorithm for Optimal Password Protection”** introduces **Cost Asymmetric Secure Hash**, a randomized key-stretching mechanism that minimizes the fraction of passwords cracked by a rational offline attacker subject to a bound on amortized authentication cost [1509.00239]. That mechanism concerns password verification economics and Stackelberg optimization rather than signal identification.

A second nearby but distinct line of work is **“\(\varepsilon\)-Almost collision-flat universal hash functions and mosaics of designs”**, which studies seeded hash families \(f:\mathcal{X}\times\mathcal{S}\to\mathcal{A}\) satisfying exact per-input output flatness and a per-output collision bound
$$
|\{ s \in \mathcal{S} : f(x,s)=f(x',s)=a \}| \le \varepsilon \frac{|\mathcal{S}|}{|\mathcal{A}|}
$$
for all distinct \(x,x'\) and every \(a\) [2306.04583]. This is a formal notion of collision control in universal hashing, but it is not the model used by the signal-processing CASH architecture.

A third relevant background is content-addressed storage. **“Collision and Preimage Resistance of the Centera Content Address”** analyzes naming schemes that combine MD5, SHA-256, random fields, timestamps, counters, and format bits to reduce practical collision risk in content addresses [1306.6020]. A plausible implication is that the phrase “collision-alleviated” in the signal model resonates with a broader literature in which collisions are mitigated structurally rather than only by increasing output length. In the SEI setting, however, the operative mechanism is not a cryptographic proof of collision resistance; it is a learned binary representation augmented by a seen/novel partition bit.

These distinctions matter because “collision” has different technical meanings across the three domains. In password hashing, the critical issue is the economics of offline guessing. In universal hashing, it is a combinatorial bound on seeded hash-family behavior. In CASH for OSEI, it is the event that distinct emitters map to the same learned binary code and are therefore merged at inference time.

## 6. Operational considerations, limitations, and prospective directions

CASH is designed so that per-sample online computation consists of one forward pass through a CVCNN, small multilayer perceptrons, and a hash-table query. The reported memory burden of the online component is limited to a code-and-identity table; for code lengths such as 13 bits after appending the indicator, the storage requirement per discovered emitter is negligible.

Several hyperparameters have a direct effect on system behavior. Hash length \(F\) controls the collision-capacity tradeoff: if \(F\) is too small, collisions increase; if \(F\) is too large, generalization to novel emitters degrades and overfitting becomes more likely. The empirical discussion reports that moderate lengths perform best, with 12 bits used for GZSL and 5 bits for FSL on the stated datasets. The ARPL radius \(R\), adversarial weight \(\lambda\), and threshold confidence \(\gamma\) influence the seen/novel separation, while \(\alpha\) and \(\beta\) balance representation learning against identifier learning and hash regularization.

The reported limitations are system-level rather than purely algorithmic. Performance under **extremely low SNR** or **rapidly varying channels** is not deeply analyzed. Scaling to **very large numbers of emitters** may require longer codes or more advanced code design. The encoder and hasher are fixed after training, so there is **no online adaptation** beyond growth of the hash table. The approach also assumes a relatively stable mapping from emitter hardware characteristics to RFF-derived hash codes; substantial hardware changes or drastic channel shifts could invalidate previously learned codes without retraining.

The paper explicitly points to future work that would connect the SEI formulation to **cryptography/information theory** for improved collision control and principled hash-length selection [2509.23807]. This suggests a possible convergence between learned signal hashing and more formal collision analyses, but within the reported system CASH remains a supervised, end-to-end SEI model whose primary contribution is to turn online emitter identification into hash-code indexing with an explicit mechanism for reducing cross-emitter code collisions.

Source: https://www.emergentmind.com/topics/collision-alleviated-signal-hash-cash