Chaos-Based Encryption: Mechanisms & Applications

• Chaos-based encryption is a family of cryptographic methods that exploit deterministic yet unpredictable chaotic dynamics to achieve confusion and diffusion.
• Techniques include discrete maps like the logistic and Arnold’s cat map, continuous systems such as Lorenz and Chua circuits, and hardware implementations like synchronized lasers.
• Evaluation relies on metrics like NPCR, UACI, and entropy, while challenges include finite-precision issues, cryptanalysis vulnerabilities, and synchronization limitations.

Chaos-based encryption (CBE) denotes a family of cryptographic constructions in which deterministic but aperiodic dynamics are used to realize confusion and diffusion, typically by permuting positions and by modifying symbols or pixels with chaotic keystreams. In the current literature, CBE spans software ciphers built from discrete maps, continuous-time dynamical systems discretized into keystreams, synchronized analog circuits and lasers, and even physical-layer devices based on wave-chaotic cavities (Rashid et al., 2024, Nazarenko et al., 2022, Park et al., 2021).

1. Dynamical basis

Digital CBE is commonly formulated through an iterated map

$x_{k+1}=g(\mu,x_k),$

with orbit $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$, where $\mu$ denotes control parameters and $x_0$ the initial state (0811.1859). In that setting, security is tied not only to the existence of chaos, but also to the observable dynamical properties of the implemented system: bifurcation structure, periodic windows, return maps, invariant measures, Lyapunov exponents, symbolic dynamics, and finite-precision effects all become cryptanalytically relevant if they leak through ciphertext or keystream statistics (0811.1859).

The principal source classes are discrete maps, continuous-time systems, and higher-dimensional or hyperchaotic dynamics. Surveyed discrete maps include the logistic map $x_{n+1}=r x_n(1-x_n)$, tent map, sine map, Chebyshev map, Arnold’s cat map for position permutation, and numerous hybrids such as logistic–sine or modular tent–logistic cascades (Rashid et al., 2024). Continuous systems widely used for sampled keystream generation include Lorenz, Chen, Rössler, and Chua dynamics; for example, the Chua circuit is modeled by

$$\frac{dv_1}{dt}=\frac{1}{C_1}\big(G(v_2-v_1)-g(v_1)\big),\quad \frac{dv_2}{dt}=\frac{1}{C_2}\big(G(v_1-v_2)+i_L\big),\quad \frac{di_L}{dt}=-\frac{1}{L}v_2,$$

with a piecewise-linear nonlinearity $g(v)$ realized by the Chua diode network (Nazarenko et al., 2022).

A recurring dynamical criterion is the Lyapunov spectrum. The general review literature treats a positive largest Lyapunov exponent as a practical indicator of chaotic behavior, while multiple positive exponents characterize hyperchaos (0811.1859, Rashid et al., 2024). In the VCSEL-based optical system, two exponents become positive as the injection current increases, and the authors explicitly interpret this as hyperchaos in the spin-flip model of polarization dynamics (Roy et al., 2017). This suggests why higher-dimensional sources are repeatedly favored in CBE: they enlarge state complexity, diversify key-dependent trajectories, and increase the difficulty of simple low-dimensional reconstruction, although this implication does not by itself constitute a proof of cryptographic security.

2. Cryptographic mechanisms and design patterns

At the algorithmic level, CBE typically separates into confusion and diffusion. Confusion permutes pixel coordinates, bit-planes, rows, columns, or blocks so that local adjacency is destroyed; diffusion then alters symbol values so that a small plaintext or key change propagates widely through the ciphertext (Rashid et al., 2024). In stream-oriented realizations, the dominant algebra is XOR:

$$c(t)=m(t)\oplus k(t),\qquad \hat m(t)=c(t)\oplus \hat k(t),$$

where $k(t)$ and $\hat k(t)$ are chaotic keystreams generated at the transmitter and receiver (Nazarenko et al., 2022).

The simplest and most common image-oriented pattern is permutation followed by substitution or XOR-based diffusion. Representative constructions include Arnold-cat-map shuffling plus Chen-system diffusion, bit-plane scrambling plus chaotic masking, chained forward/reverse diffusion passes, and blockwise permutations driven by chaotic indices (Mukherjee et al., 2018, Rashid et al., 2024). More elaborate variants combine chaos with quasigroup substitution tables, dynamic S-boxes, cellular automata, DNA/RNA coding rules, or Feistel-like structures. The quasigroup-based scheme of Sufi et al. replaces pixel values through Latin-square lookup operations and then applies standard-map-driven row and column permutations, explicitly using the quasigroup itself as part of the secret key (Patidar et al., 2017). The Feistel-structured image/video design of Naskar et al. uses a logistic-map byte keystream together with a Permutation–Inverse Permutation transform over 8-pixel, 64-bit blocks, so that bit positions are scrambled before XOR and restored afterward (Patnayak et al., 2020).

A recurrent design goal is to improve the statistical quality of chaotic outputs before they are used cryptographically. One route is post-processing or whitening. The Lorenz–Chen voice cipher extracts fractional parts, thresholds them into six binary streams, mixes them by XOR, and selects the final bit through a dynamic 4-to-1 multiplexer (Ahmad et al., 2014). Another route is hybrid PRNG design: the unit-transform coupling framework of Li et al. produces enhanced one-dimensional chaotic maps and then XORs their integer outputs with a linear congruential generator in the CBPRNG, for which the paper gives a uniformity argument and reports full passage of Rabbit, Alphabit, and NIST SP 800-22 batteries (Hu et al., 2019). A more recent route is data-identified dynamics: the SINDy-PI scheme learns a discrete chaotic map directly from trajectories and then feeds its outputs into a sorting-based permutation stage and multi-round modulo-256 diffusion (Lia et al., 20 May 2026).

3. Hardware and physical realizations

CBE is not restricted to software simulation. A direct hardware demonstrator based on synchronized Chua circuits places one chaotic circuit in the transmitter and one in the receiver, coupled by a dedicated bidirectional synchronization line while ciphertext travels on a separate line (Nazarenko et al., 2022). The transmitter digitizes its chaotic voltage and XORs it with a plaintext square wave specified in the demonstrator as 6 kHz, 2.5 Vpp, with +1.25 V DC offset and 50% duty cycle; the receiver digitizes its synchronized chaotic signal, XORs again to recover the message, and applies a first-order low-pass filter with $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$0 and $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$1 (Nazarenko et al., 2022). The reported phase portraits show a Chua attractor at the transmitter and a near-diagonal synchronization portrait between transmitter and receiver states, which the authors interpret as near-perfect synchronization sufficient for stream decryption (Nazarenko et al., 2022).

Optical implementations use synchronized lasers rather than electronic oscillators. In the VCSEL system, a bidirectionally coupled transmitter–receiver pair is modeled by the spin-flip equations for right- and left-circularly polarized fields and carrier populations, with reported synchronization error reaching the order of $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$2 under a specific parameter setting $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$3 and $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$4 (Roy et al., 2017). That system supports direct image encryption and reports strong statistical results at the image level, but its architectural significance lies in demonstrating that synchronization-based CBE can be implemented in nonlinear optical media rather than solely in digital logic or analog electronics (Roy et al., 2017).

The most radical extension is physical-layer CBE through wave chaos. The wave-chaos device uses a compact quasi-2D disordered cavity as a physical keyed transform: a transmitted waveform is temporally distorted by the cavity impulse response, and successful recovery requires a receiver cavity with effectively identical geometry and port arrangement (Park et al., 2021). In the reported wireless experiment, the matched cavity pair exhibits impulse-response correlation $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$5, whereas two mismatched cavities give $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$6 and $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$7; the demonstration operates with UWB pulses from 3–10 GHz, symbol period $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$8, and bit rate $\gamma(\mu,x_0)=\{x_0,x_1,\ldots\}$9 (Park et al., 2021). Unlike software keystream generation, this realization makes the physical transfer function itself part of the key material.

4. Evaluation methodology

CBE papers overwhelmingly evaluate security through image-statistical and sequence-randomness metrics. The standard image metrics are NPCR,

$\mu$0

UACI,

$\mu$1

information entropy

$\mu$2

and adjacent-pixel correlation coefficients in horizontal, vertical, and diagonal directions (Rashid et al., 2024). Histogram flatness, chi-square tests, PSNR/MSE under data loss, and NIST SP 800-22 are also repeatedly used (Rashid et al., 2024, Ahmad et al., 2014).

Reported results vary widely across schemes. The VCSEL image cipher gives NPCR values of 99.73%, 99.78%, and 99.74% for the R, G, and B channels, UACI values of 33.23%, 33.27%, and 33.14%, and cipher entropies around 7.99 bits per channel, with adjacent-pixel correlations near zero in all directions (Roy et al., 2017). The hearing-aid-oriented TD-ERCS/NCA system reports NPCR = 98.5537, UACI = 32.8331, entropy = 7.9681, encrypted-direction correlations DCC = 0.0033, HCC = -0.0018, VCC = -0.0018, and mean single-image runtime 0.00258 s on a Ryzen 5 4600H / GTX 1650Ti platform (Shah et al., 2022). CellSecure reports grayscale entropies 7.9958 and 7.9944 for Baboon and Cameraman, with ciphertext correlations such as -0.0048808, -0.0010709, and 0.0047433 for Baboon and 0.0022096, -0.00446, and 0.0043268 for Cameraman (Ali et al., 2023).

Recent data-driven work explicitly targets the same benchmark region. The SINDy-PI map-identification scheme reports entropy 7.9976, ciphertext chi-square 212.7969 with $\mu$3, average NPCR 99.6556%, average UACI 33.5824%, and decryption failure under key perturbations as small as $\mu$4 (Lia et al., 20 May 2026). In voice encryption, the mixed Lorenz–Chen generator passes the listed NIST SP 800-22 tests on 100,000 keystream bits, while the maximum non-zero-shift autocorrelation of the encrypted voice falls to 0.0152536 from 0.8707092 for the original signal (Ahmad et al., 2014). At the same time, not all CBE papers report the same depth of evaluation: the Chua-circuit hardware demonstrator shows successful synchronization and visually perfect decryption, but does not provide NIST SP 800-22, entropy-per-bit, BER, or latency measurements (Nazarenko et al., 2022). This unevenness in evaluation methodology is itself a defining feature of the field.

5. Cryptanalysis, reliability, and recurring weaknesses

A major theme in the literature is that dynamical complexity does not automatically imply cryptographic robustness. The general cryptanalytic framework for digital chaos-based cryptography treats return maps, bifurcation diagrams, Lyapunov exponents, invariant densities, entropy profiles, and symbolic dynamics as attack surfaces; if ciphertext or keystream leaks enough of the underlying dynamics, parameters or equivalent keys may be recoverable (0811.1859). This concern is not merely theoretical. Several image ciphers built from low-dimensional maps and separable permutation–diffusion stages have been broken under chosen-plaintext conditions.

One line of work attacks insufficient diffusion. The revision of a cat-map-based image cipher shows that a single round does not deliver acceptable avalanche or histogram uniformity, and that at least six iterations are needed to reach approximately 50% avalanche across tested image sizes; the same paper also argues that full-image avalanche induces severe error propagation, recommending block-level avalanche instead (Noura et al., 2017). Another line attacks separability and keystream reuse. A Chen-system-plus-Arnold-map design can be completely broken with two chosen queries because the keystream is plaintext-independent and the permutation stage can be learned separately (Mukherjee et al., 2018). The older MCKBA scheme, driven by a logistic map and simple modular-addition/XOR/XNOR structure, is broken with four chosen plain-images, and the paper additionally reports ineffective key MSBs and poor diffusion (Li et al., 2010). A color image cipher based on logistic-map sequences, two permutation stages, and one substitution stage can be decomposed even more finely: its substitution part is recoverable with two chosen constant images, and its permutations with only $\mu$5 chosen plain-images (Li et al., 2012).

More recent work extends cryptanalysis beyond classical algebraic or differential models. A deep-learning attack on the Guan 2005 Arnold-cat/Chen construction learns a direct ciphertext-to-plaintext inversion without recovering the key; on MNIST, the reconstructed images remain sufficiently faithful that a LeNet-5 classifier achieves 97.87% accuracy in the static-key case and 92.04% in a dynamic-$\mu$6 scenario (He et al., 2019). This suggests that some chaotic image ciphers may leak learnable structure even when conventional histogram and entropy tests appear favorable.

Reliability under real computation is a separate but equally important issue. When the same cubic map with the same initial condition was iterated on two IEEE 754-compliant devices, the resulting pseudo-orbits diverged rapidly enough that the derived keystreams differed and cross-device decryption failed; the reported normalized key entropies were 0.97206928 and 0.97283489 on the two platforms (Santos et al., 2019). A mitigation based on scaling the update by $\mu$7 reduces the effective Lyapunov growth and yields partially readable decryption, but does not restore exact reproducibility (Santos et al., 2019). This makes finite precision, arithmetic non-determinism, and platform dependence central concerns for any CBE design that depends on synchronized numerical iteration rather than explicit authenticated key exchange or discrete-state generation.

6. Applications, hybrids, and current directions

Despite these criticisms, CBE remains active because it fits application domains where lightweight arithmetic, structure-aware scrambling, or physical synchronization are attractive. The literature surveyed for satellite communications presents CBE as a candidate privacy mechanism for image-heavy LEO/MEO/GEO links and emphasizes low latency, limited onboard resources, and the usefulness of blockwise or hybrid designs under CCSDS constraints (Rashid et al., 2024). Other reported domains include IIoT and CPS image protection through logistic–sine maps and Rule-30 cellular automata (Ali et al., 2023), multimodal hearing aids with TD-ERCS/NCA and SHA-512-based seeding (Shah et al., 2022), voice masking with mixed Lorenz–Chen keystreams (Ahmad et al., 2014), real-time image and video encryption via logistic keystreams and PIP transforms (Patnayak et al., 2020), synchronized optical links (Roy et al., 2017), low-power embedded links with Chua circuits (Nazarenko et al., 2022), and wireless wave transmission using matched chaotic cavities (Park et al., 2021).

A visible trend is hybridization. The satellite review repeatedly recommends schemes that combine chaos-based permutation or block selection with standard symmetric primitives such as AES-CTR or AES-GCM, in order to preserve authenticated-encryption properties while still exploiting image-aware scrambling and low-overhead chaotic indexing (Rashid et al., 2024). Other hybrids expand the key space through non-chaotic combinatorial structures, as in quasigroup substitution tables (Patidar et al., 2017), or improve sequence quality through chaos-plus-LCG whitening, as in CBPRNG (Hu et al., 2019).

A second trend is methodological diversification. The UT-CCS framework couples one-dimensional maps through unit transforms chosen to enlarge Lyapunov exponents across the full parameter range $\mu$8, and its Type III variants empirically dominate simpler seed maps in sample entropy and LE profiles (Hu et al., 2019). The SINDy-PI work moves in a different direction by making the map itself data-identified rather than fixed a priori; with the same initial conditions, maps learned from slightly different training data produce ciphertexts with NPCR $\mu$9 and UACI $x_0$0, effectively treating model identification as an additional variability layer (Lia et al., 20 May 2026). This suggests a shift from classical hand-chosen maps toward learned or system-specific dynamics.

The open problems remain consistent across otherwise very different papers. Repeatedly cited needs include formal or at least more systematic cryptanalysis, comprehensive statistical test batteries, quantified key-space and mismatch analysis, side-channel and synchronization-line leakage studies, finite-precision and reproducibility analysis, and better integration with standard authenticated cryptography (Nazarenko et al., 2022, Rashid et al., 2024, Santos et al., 2019). In that sense, CBE is best understood not as a single cipher family with settled security doctrine, but as a heterogeneous research area at the intersection of nonlinear dynamics, signal processing, hardware synchronization, and applied cryptography.