Proxy Target Networks in Computing

• Proxy Target Networks are frameworks where proxy entities mediate target functions, enabling efficient neural architecture search, dynamic network overlays, and stable neuromorphic updates.
• They leverage latency monotonicity and lightweight regression adaptation to drastically reduce profiling costs in hardware-aware systems and optimize traffic routing in communication architectures.
• Future challenges include handling non-monotonic conditions, ensuring accurate telemetry in proxies, and extending applications to complex search tasks and resource-constrained environments.

Proxy Target Networks refers to a collection of models, methods, or architectural patterns where proxy entities (networks, devices, or nodes) mediate, facilitate, or serve as surrogates for target functions in diverse computational, algorithmic, and networking contexts. These structures appear in hardware-aware machine learning, networking, neuromorphic computing, and distributed systems, providing scalability, compatibility, resource protection, policy enforcement, or algorithmic stability. The unifying theme is the use of an intermediate “proxy”—whether a device, network, or function—that either estimates, emulates, negotiates, or relays interactions and optimization flows intended for a more resource-constrained, non-differentiable, or otherwise inaccessible “target”.

1. Proxy Target Networks in Hardware-Aware Neural Architecture Search

In hardware-aware neural architecture search (NAS), the dominant cost is constructing device-specific latency predictors by exhaustive profiling. The proxy target network framework efficiently addresses this by exploiting the empirical latency monotonicity between architecture rankings on a “proxy” device and a set of “target” devices. Specifically, if the latency orderings ℓ(x; d) for architectures x are nearly monotonic between devices (i.e., Spearman's ρ ≈1), Pareto-optimal architectures discovered via NAS on the proxy device d₀ can be transferred directly to any target device d. This substantially reduces the cost from O(n · #samples) to O(1 · #samples) per new device (Lu et al., 2021).

When monotonicity is weaker (ρ<0.9), a lightweight proxy adaptation step is formulated: for M ≲ 50 architectures measured on the target device, a regression-based correction recalibrates the proxy’s linear latency model:

$$\min_{a,b} \sum_{i=1}^{M} \left[(a \cdot 1 + b)^{\mathsf{T}}(w \odot x_i) - y_i \right]^2 + \lambda \|b\|_1$$

Here $w$ are the original proxy weights, $x_i$ are architecture encodings, and the adapted predictor $\hat \ell(d_0 \to d)(x)$ then recovers the true Pareto front within ≲1% of full per-device NAS. This approach has been demonstrated across diverse hardware, including mobile, GPU, FPGA, and VPU platforms, enabling scalable deployment of optimized CNNs and resource-efficient per-device specialization without exhaustive re-profiling (Lu et al., 2021).

2. Overlay Proxy Target Networks in Communication Architectures

Proxy target networks also describe logical overlays of programmable proxies that intercept, forward, and process traffic on behalf of applications or legacy protocols. In the Hermes architecture (Farkiani et al., 2024), proxies are partitioned into dependent proxies (DPs) located at the edge (e.g., within end-hosts or services) and standalone proxies (SaPs) forming the overlay backbone. Traffic is routed through chains of these proxies using HTTP-embedded headers carrying routing semantics, creating dynamically reconfigurable “proxy target networks”.

The architecture supports a wide range of tunneling and encapsulation mechanisms, including TCP-over-HTTP (CONNECT), UDP-over-HTTP (CONNECT-UDP), and IP-over-HTTP (CONNECT-IP per MASQUE standard). Telemetry is streamed from each proxy, enabling centralized controllers to optimize path selection against metrics such as end-to-end latency:

$L(P) = \sum_{i=0}^{n-1} \left[l_i + d_i \right]$

and throughput/bottleneck analysis. Reliability, policy-based pathing, and automatic configuration adaptation allow Hermes overlays to transparently support legacy applications, deliver high reliability over lossy multihop links, and serve as a substrate for future architectures like Named Data Networking. Experimental results confirm end-to-end benefits in video streaming, VPN replacement, and intermittent network recovery, with fine-grained policy and traffic engineering based on proxy header semantics (Farkiani et al., 2024).

3. Proxy Target Networks in Neuromorphic and Reinforcement Learning Algorithms

Proxy target frameworks provide algorithmic bridges in neuromorphic RL by employing differentiable proxy networks for stable target updates in the presence of non-differentiable spiking neurons. In standard continuous-control Actor-Critic methods (e.g., TD3), the “soft” target update mechanism is ill-posed for Spiking Neural Networks (SNNs) due to the non-differentiable nature of spikes. The Proxy Target approach substitutes the SNN target actor with a small, fully differentiable proxy ANN (“proxy target network”) during training:

• The proxy network tracks the online spiking actor via a surrogate L₂ loss.
• The Bellman backup for the critic uses outputs from the proxy, ensuring smooth evolution and reversibility of target policies otherwise subject to abrupt SNN-induced discontinuities.

Formally, the target update is:

$$\pi_\phi'^\text{Proxy}(s) \leftarrow (1-\tau)\,\pi_\phi'^\text{Proxy}(s) + \tau\, \pi_\phi^\text{SNN}(s)$$

implemented via stochastic gradient descent on

$$L_\text{proxy} = \frac{1}{N} \sum_{i=1}^N \| \pi_\phi'^\text{Proxy}(s_i) - \pi_\phi^\text{SNN}(s_i)\|_2^2$$

This enables energy-efficient SNN policies to achieve or surpass ANN baseline performance without sacrificing training stability; at deployment, the proxy is discarded and only the SNN is used for inference, preserving the efficiency and low-latency of neuromorphic hardware (Xu et al., 30 May 2025).

4. Guard Proxies and Proxy Target Networks in Constrained RESTful Environments

In CoRE (Constrained RESTful Environments), proxy target networks are realized by deploying guard proxies at ingress points to resource-constrained subnets (e.g., IoT CoAP/OSCORE networks). These guard proxies serve as application-aware firewalls, implementing roles such as request mediation, rate limiting, and token-based authentication:

• A “Proxy Target Network” consists of the devices whose access is mediated entirely by one or more guard proxies.
• Discovery of such proxies occurs via rendezvous services (DNS-SD, Resource Directory) or multicast queries, with probabilistic models governing discovery reliability and announcement overhead.

Three intrusiveness levels are distinguished:

Proxy Type	Functionality	Resource Impact
Transparent	UDP/IP forwarding only	Minimal RAM/CPU
App-aware	CoAP/OSCORE, throttling	More RAM/CPU
Full Guard	Token negotiation, tunnels	Highest RAM/CPU

Guard proxies provide quantifiable protection against resource exhaustion (energy denial, bandwidth attack), with their effectiveness directly modeled via queueing and energy budget equations. The framework extends to composite deployments, dynamic adaptation, and balancing between legitimate traffic completion and attack mitigation (Amsüss, 2022).

5. Proxy Target as Analytical Proxy for Complex Search on Networks-within-Networks

In the stochastic processes literature, “proxy” targets or fluxes serve as analytical stand-ins for mean-first passage times (MFPTs) in hierarchical, multiplex network search problems. A search process may involve an external source, intermediate spatial network, and a local internal-state network. The steady-state flux $J_\infty$ to a specific target node (potentially in a specific internal state) is universally related to the inverse MFPT:

$J_\infty \approx \frac{1}{\mathrm{MFPT}}$

This “proxy” encapsulates search efficiency, with closed-form expressions derived via matrix inverses of network Laplacians and absorption rates. This proxy target methodology unifies optimal design and analysis in biological search (e.g., protein–DNA binding), epidemiology (infection pathways), and transportation networks, providing a generalizable scaling law for target search efficiency in systems exhibiting networks-within-networks architecture (Hedström et al., 2024).

6. Limitations, Open Challenges, and Future Directions

The proxy target network paradigm, while widely applicable, imposes certain constraints and leads to domain-specific challenges:

• In hardware-aware NAS, the main limitation is the breakdown of latency monotonicity when the search space spans both memory-bound and compute-bound regimes, or when input/dataset modifications alter hardware behavior, necessitating re-adaptation (Lu et al., 2021).
• Overlay network proxies depend on reliable telemetry and correct policy configuration; dynamic or adversarial conditions may require rapid reconfiguration and ongoing monitoring for performance regression or security exposure (Farkiani et al., 2024).
• For SNN/ANN proxy targets in RL, extended credit assignment (e.g., eligibility traces) and generalization to stochastic or on-policy methods remain open research directions (Xu et al., 30 May 2025).
• In CoRE environments, provisioning resource-limited proxies and managing the trade-off between resource protection and service availability in large, unmanaged deployments poses ongoing architectural and standardization questions (Amsüss, 2022).
• Analytical proxy target models rely on assumptions of stationary, Markovian dynamics and dilute searcher concentration; departures from these regimes may require more nuanced or simulation-based proxies (Hedström et al., 2024).

A plausible implication is that, as systems become more heterogeneous and operationally dynamic, proxy target frameworks—physical, logical, or analytical—will increasingly underpin strategies for scalable optimization, adaptive control, and secure mediation across networked computational ecosystems.