diff --git a/include/aidge/backend/cpu/operator/HeavisideImpl.hpp b/include/aidge/backend/cpu/operator/HeavisideImpl.hpp
index 7a3ba9add1e98580c51a8416adc0d1feb5e1317a..877fa2a9c1fbf126fee5d1f3ce4db2db808cbc92 100644
--- a/include/aidge/backend/cpu/operator/HeavisideImpl.hpp
+++ b/include/aidge/backend/cpu/operator/HeavisideImpl.hpp
@@ -23,7 +23,7 @@ namespace Aidge {
 using HeavisideImplCpu =
     OperatorImpl_cpu<Heaviside_Op,
                      void(std::size_t, const void *, void *, const float),
-                     void(const float, std::size_t, const void *, void *)>;
+                     void(std::size_t, const void *, const void *, void *)>;
 
 // Implementation entry point registration for operator Heaviside
 REGISTRAR(Heaviside_Op, "cpu", HeavisideImplCpu::create);
diff --git a/include/aidge/backend/cpu/operator/HeavisideImpl_kernels.hpp b/include/aidge/backend/cpu/operator/HeavisideImpl_kernels.hpp
index 0bbbddee1040224fd9d0c2658f97c57e3956ca48..7fc0eb0a86210d896ad0ccec9b88150670f4e328 100644
--- a/include/aidge/backend/cpu/operator/HeavisideImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/HeavisideImpl_kernels.hpp
@@ -48,7 +48,8 @@ void HeavisideImplCpuBackwardKernel(std::size_t inputLength,
 
     for (size_t i = 0; i < inputLength; ++i) {
         // dx = dy * (1/PI) * (1 / (1 + (PI * x)^2))
-        grad_input[i] = (1 / M_PI) * grad_output[i] * static_cast<O>(1.0 / (1.0 + output[i] * output[i]));
+        // grad_input[i] = (1 / M_PI) * grad_output[i] * static_cast<O>(1.0 / (1.0 + (output[i] * output[i]) * (M_PI * M_PI)));
+        grad_input[i] = grad_output[i] * static_cast<O>(1.0 / (1.0 + (output[i] * output[i]) * (M_PI * M_PI)));
     }
 }
 
@@ -57,7 +58,7 @@ REGISTRAR(HeavisideImplCpu,
           {DataType::Float32},
           {ProdConso::inPlaceModel,
            Aidge::HeavisideImplCpuForwardKernel<float, float>,
-           nullptr});
+           Aidge::HeavisideImplCpuBackwardKernel<float,float,float>});
 } // namespace Aidge
 
 #endif // AIDGE_CPU_OPERATOR_HEAVISIDEIMPL_KERNELS_H__H_
diff --git a/src/operator/HeavisideImpl.cpp b/src/operator/HeavisideImpl.cpp
index 2ead2978ef1c17ec38e2d12df6b1c78f6c894964..5bf77f87072005896b87f009ebe78711606f78ff 100644
--- a/src/operator/HeavisideImpl.cpp
+++ b/src/operator/HeavisideImpl.cpp
@@ -34,21 +34,22 @@ template <> void Aidge::HeavisideImplCpu::forward() {
 
 template <> 
 void Aidge::HeavisideImplCpu::backward() {
-    AIDGE_THROW_OR_ABORT(std::runtime_error, "Heaviside backward not implemented yet");
 
     // TODO: The following lines are assuming that the surrogate gradient is Atan
     // remove that assumption by providing an attribute to Heaviside, 
     // allowing to choose between different surrogate gradients.
     
-    // const Heavisde_Op& op_ = dynamic_cast<const Heavisie_Op &>(mOp);
+    const Heaviside_Op& op_ = dynamic_cast<const Heaviside_Op &>(mOp);
 
+    const auto impl = Registrar<HeavisideImplCpu>::create(getBestMatch(getRequiredSpec()));
 
+    auto gra_int0 = op_.getInput(0)->grad();
+    auto gra_out0 = op_.getOutput(0)->grad();
 
-    // ! backward of hs = forward of atan
-    //const auto impl = Registrar<HeavisideImplCpu>::create(getBestMatch(getRequiredSpec()));
-    // std::shared_ptr<Tensor> in0 = op_.getInput(0);
-    // std::shared_ptr<Tensor> out0 = op_.getOutput(0);
-  
-    //impl.forward()
+    std::shared_ptr<Tensor> in0 = op_.getInput(0);
+    std::shared_ptr<Tensor> out0 = op_.getOutput(0);
+    AIDGE_ASSERT(out0, "missing output #0 for current {} operator", op_.type());
+
+    impl.backward(gra_int0->size(), getCPUPtr(in0), getCPUPtr(gra_out0), getCPUPtr(gra_int0));
 }
 
diff --git a/unit_tests/operator/Test_HeavisideImpl.cpp b/unit_tests/operator/Test_HeavisideImpl.cpp
index a0142513d20d4df474d69582e89c8648a06fa340..515d6802d56f2c4a90ecf0bc2d86f2e9a727aeed 100644
--- a/unit_tests/operator/Test_HeavisideImpl.cpp
+++ b/unit_tests/operator/Test_HeavisideImpl.cpp
@@ -12,6 +12,7 @@
 #include "aidge/backend/cpu/operator/HeavisideImpl_kernels.hpp"
 
 #include <aidge/operator/Memorize.hpp>
+#include <aidge/utils/Types.h>
 #include <memory>
 #include <cstdlib>
 #include <random>
@@ -100,10 +101,29 @@ TEST_CASE("[cpu/operator] Heaviside(forward)", "[Heaviside][CPU]") {
 }
 
 TEST_CASE("[cpu/operator] Heaviside(backward)", "[Heaviside][CPU]") {
+    auto hs = Heaviside(1.0f);
+    auto op = std::static_pointer_cast<OperatorTensor>(hs->getOperator());
+    op->setDataType(DataType::Float32);
+    op->setBackend("cpu");
 
-    auto add = Add();
-    auto mem = Memorize(2);
-    auto hs = Heaviside(1);
-}
+    auto input = Tensor(Array1D<float, 3>({1.0, -1.0, 1.0}));
+    input.setDataType(DataType::Float32);
+    input.setBackend("cpu");
+
+    auto grad = Tensor(Array1D<float, 3>({1.0, 1.0, 1.0}));
+    grad.setDataType(DataType::Float32);
+    grad.setBackend("cpu");
+
+    op->setInput(IOIndex_t(0), std::make_shared<Tensor>(input));
+    op->forward();
 
+    Log::info("Output : ");
+    op->getOutput(0)->print();
+
+    op->getOutput(0)->setGrad(std::make_shared<Tensor>(grad));
+    op->backward();
+
+    Log::info("Gradient : ");
+    op->getInput(0)->grad()->print();
+}
 }