diff --git a/Dockerfile b/Dockerfile
index 4907780e..3dea5d8c 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,8 +1,9 @@
 FROM python:3.8 AS xaitk_base
 
 # install poetry and configure to install to pip environment
-RUN pip install poetry \
-&& poetry config virtualenvs.create false
+ENV PATH=/root/.local/bin:${PATH}
+RUN (curl -sSL 'https://install.python-poetry.org' | python3 -) \
+ && poetry config virtualenvs.create false
 
 WORKDIR /xaitk
 
@@ -11,5 +12,9 @@ COPY poetry.lock pyproject.toml /xaitk/
 RUN poetry install --no-root
 
 # install package
-COPY . /xaitk/
+COPY docs /xaitk/docs/
+COPY scripts /xaitk/scripts/
+COPY tests /xaitk/tests/
+COPY xaitk_saliency /xaitk/xaitk_saliency
+COPY .flake8 .mypy.ini CONTRIBUTING.md LICENSE.txt README.md /xaitk/
 RUN poetry install
diff --git a/docs/release_notes/pending_release.rst b/docs/release_notes/pending_release.rst
index 4b1e03e6..2990e286 100644
--- a/docs/release_notes/pending_release.rst
+++ b/docs/release_notes/pending_release.rst
@@ -4,9 +4,20 @@ Pending Release Notes
 Updates / New Features
 ----------------------
 
+Docker
+
+* Update Dockerfile install of poetry, and make separate & specific directory
+  copies.
+
 Utils
 
 * Updated logging format of occlusion masking benchmark utility.
 
 Fixes
 -----
+
+Tests
+
+* Update various floating point equality comparisons to not use exact match.
+
+* Fix random number usage from numpy to use `np.random.default_rng`.
diff --git a/poetry.lock b/poetry.lock
index cc9e2147..76176e16 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.5.1 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.7.1 and should not be changed by hand.
 
 [[package]]
 name = "affine"
@@ -1371,6 +1371,7 @@ optional = true
 python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*, !=3.6.*"
 files = [
     {file = "jsonpointer-2.4-py2.py3-none-any.whl", hash = "sha256:15d51bba20eea3165644553647711d150376234112651b4f1811022aecad7d7a"},
+    {file = "jsonpointer-2.4.tar.gz", hash = "sha256:585cee82b70211fa9e6043b7bb89db6e1aa49524340dde8ad6b63206ea689d88"},
 ]
 
 [[package]]
@@ -2553,7 +2554,7 @@ files = [
 [package.dependencies]
 numpy = [
     {version = ">=1.20.3", markers = "python_version < \"3.10\""},
-    {version = ">=1.21.0", markers = "python_version >= \"3.10\""},
+    {version = ">=1.21.0", markers = "python_version >= \"3.10\" and python_version < \"3.11\""},
     {version = ">=1.23.2", markers = "python_version >= \"3.11\""},
 ]
 python-dateutil = ">=2.8.2"
@@ -2634,6 +2635,7 @@ description = "parse() is the opposite of format()"
 optional = true
 python-versions = "*"
 files = [
+    {file = "parse-1.19.0-py2.py3-none-any.whl", hash = "sha256:6ce007645384a91150cb7cd7c8a9db2559e273c2e2542b508cd1e342508c2601"},
     {file = "parse-1.19.0.tar.gz", hash = "sha256:9ff82852bcb65d139813e2a5197627a94966245c897796760a3a2a8eb66f020b"},
 ]
 
@@ -3478,6 +3480,7 @@ files = [
     {file = "scikit_image-0.21.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8ef5d8d1099317b7b315b530348cbfa68ab8ce32459de3c074d204166951025c"},
     {file = "scikit_image-0.21.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:78b1e96c59cab640ca5c5b22c501524cfaf34cbe0cb51ba73bd9a9ede3fb6e1d"},
     {file = "scikit_image-0.21.0-cp39-cp39-win_amd64.whl", hash = "sha256:9cffcddd2a5594c0a06de2ae3e1e25d662745a26f94fda31520593669677c010"},
+    {file = "scikit_image-0.21.0.tar.gz", hash = "sha256:b33e823c54e6f11873ea390ee49ef832b82b9f70752c8759efd09d5a4e3d87f0"},
 ]
 
 [package.dependencies]
diff --git a/tests/impls/gen_classifier_conf_sal/test_occlusion_scoring.py b/tests/impls/gen_classifier_conf_sal/test_occlusion_scoring.py
index 228388bd..6d03b9f3 100644
--- a/tests/impls/gen_classifier_conf_sal/test_occlusion_scoring.py
+++ b/tests/impls/gen_classifier_conf_sal/test_occlusion_scoring.py
@@ -58,11 +58,11 @@ def test_1class_scores(self) -> None:
         Test basic scoring with a single class for broadcasting sanity check.
         """
         impl = OcclusionScoring()
-        np.random.seed(2)
+        rng = np.random.default_rng(2)
         # Three Perturbation masks of height and width 10px for 1 class
-        image_confs_1_class_ = np.random.rand(1)
-        pertb_confs_1_class_ = np.random.rand(3, 1)
-        mask_confs_1_class_ = np.random.randint(low=0, high=2, size=(3, 10, 10), dtype='int')
+        image_confs_1_class_ = rng.standard_normal((1))
+        pertb_confs_1_class_ = rng.standard_normal((3, 1))
+        mask_confs_1_class_ = rng.integers(low=0, high=2, size=(3, 10, 10), dtype='int')
 
         sal = impl.generate(image_confs_1_class_, pertb_confs_1_class_, mask_confs_1_class_)
         assert sal.shape == (1, 10, 10)
@@ -85,10 +85,10 @@ def test_20class_scores(self) -> None:
         Test scoring for 20 classes.
         """
         impl = OcclusionScoring()
-        np.random.seed(2)
+        rng = np.random.default_rng(2)
         # Three Perturbation masks of height and width 10px for 20 classes
-        image_confs_1_class_ = np.random.rand(20)
-        pertb_confs_1_class_ = np.random.rand(3, 20)
-        mask_confs_1_class_ = np.random.randint(low=0, high=2, size=(3, 10, 10), dtype='int')
+        image_confs_1_class_ = rng.standard_normal((20))
+        pertb_confs_1_class_ = rng.standard_normal((3, 20))
+        mask_confs_1_class_ = rng.integers(low=0, high=2, size=(3, 10, 10), dtype='int')
         sal = impl.generate(image_confs_1_class_, pertb_confs_1_class_, mask_confs_1_class_)
         assert sal.shape == (20, 10, 10)
diff --git a/tests/impls/gen_descriptor_sim_sal/test_similarity_scoring.py b/tests/impls/gen_descriptor_sim_sal/test_similarity_scoring.py
index 2288b095..9c5e707e 100644
--- a/tests/impls/gen_descriptor_sim_sal/test_similarity_scoring.py
+++ b/tests/impls/gen_descriptor_sim_sal/test_similarity_scoring.py
@@ -57,11 +57,11 @@ def test_generate_mismatch_ref_descriptors(self) -> None:
         Test that we appropriately error when the input reference descriptors
         are not the same dimensionality.
         """
-        np.random.seed(0)
-        test_ref_descr = np.random.rand(16)
-        test_query_descrs = np.random.rand(5, 15)  # Different than above
-        test_pert_descrs = np.random.rand(32, 16)
-        test_masks = np.random.rand(32, 16, 16)
+        rng = np.random.default_rng(seed=0)
+        test_ref_descr = rng.standard_normal(16)
+        test_query_descrs = rng.standard_normal((5, 15))  # Different than above
+        test_pert_descrs = rng.standard_normal((32, 16))
+        test_masks = rng.standard_normal((32, 16, 16))
 
         impl = SimilarityScoring()
 
@@ -76,11 +76,11 @@ def test_generate_mismatched_perturbed(self) -> None:
         Test that we appropriately error when the input perturbation
         descriptors and mask arrays are not equal in first-dimension length.
         """
-        np.random.seed(0)
-        test_ref_descr = np.random.rand(16)
-        test_query_descrs = np.random.rand(1, 16)
-        test_pert_descrs = np.random.rand(32, 16)
-        test_masks = np.random.rand(30, 16, 16)  # Different than above
+        rng = np.random.default_rng(seed=0)
+        test_ref_descr = rng.standard_normal(16)
+        test_query_descrs = rng.standard_normal((1, 16))
+        test_pert_descrs = rng.standard_normal((32, 16))
+        test_masks = rng.standard_normal((30, 16, 16))  # Different than above
 
         impl = SimilarityScoring()
 
@@ -96,11 +96,11 @@ def test_1_featurevec(self) -> None:
         Test basic scoring with a single feature for broadcasting sanity check.
         """
         impl = SimilarityScoring()
-        np.random.seed(2)
-        ref_feat = np.random.rand(2048)
-        query_feats = np.random.rand(2, 2048)
-        pertb_feats = np.random.rand(3, 2048)
-        pertb_mask = np.random.randint(low=0, high=2, size=(3, 10, 10), dtype='int')
+        rng = np.random.default_rng(2)
+        ref_feat = rng.standard_normal(2048)
+        query_feats = rng.standard_normal((2, 2048))
+        pertb_feats = rng.standard_normal((3, 2048))
+        pertb_mask = rng.integers(low=0, high=2, size=(3, 10, 10), dtype='int')
         sal = impl.generate(ref_feat, query_feats, pertb_feats, pertb_mask)
         assert sal.shape == (2, 10, 10)
 
@@ -123,10 +123,10 @@ def test_512_featdim(self) -> None:
         Test scoring for features of 512 dims.
         """
         impl = SimilarityScoring()
-        np.random.seed(2)
-        ref_feat = np.random.rand(512)
-        query_feats = np.random.rand(1, 512)
-        pertb_feats = np.random.rand(15, 512)
-        pertb_mask = np.random.randint(low=0, high=2, size=(15, 10, 10), dtype='int')
+        rng = np.random.default_rng(2)
+        ref_feat = rng.standard_normal(512)
+        query_feats = rng.standard_normal((1, 512))
+        pertb_feats = rng.standard_normal((15, 512))
+        pertb_mask = rng.integers(low=0, high=2, size=(15, 10, 10), dtype='int')
         sal = impl.generate(ref_feat, query_feats, pertb_feats, pertb_mask)
         assert sal.shape == (1, 10, 10)
diff --git a/tests/impls/gen_detector_prop_sal/test_drise_scoring.py b/tests/impls/gen_detector_prop_sal/test_drise_scoring.py
index f9028b45..171c858a 100644
--- a/tests/impls/gen_detector_prop_sal/test_drise_scoring.py
+++ b/tests/impls/gen_detector_prop_sal/test_drise_scoring.py
@@ -23,9 +23,10 @@ def test_shape_sanity(self) -> None:
         Test basic scoring with a single feature for broadcasting sanity check.
         """
         impl = DRISEScoring()
-        ref_dets = np.random.rand(2, 7)
-        pert_dets = np.random.rand(10, 3, 7)
-        pert_mask = np.random.randint(
+        rng = np.random.default_rng(seed=0)
+        ref_dets = rng.standard_normal((2, 7))
+        pert_dets = rng.standard_normal((10, 3, 7))
+        pert_mask = rng.integers(
             low=0, high=2, size=(10, 15, 25), dtype='int')
         sal = impl.generate(ref_dets, pert_dets, pert_mask)
         assert sal.shape == (2, 15, 25)
@@ -52,9 +53,10 @@ def test_mask_detections_size_mismatch(self) -> None:
         Test size mismatch between perturbed detections and masks.
         """
         impl = DRISEScoring()
-        ref_dets = np.random.rand(2, 7)
-        pert_dets = np.random.rand(9, 3, 7)  # ONE LESS than pert mask mat.
-        pert_mask = np.random.randint(
+        rng = np.random.default_rng(seed=0)
+        ref_dets = rng.standard_normal((2, 7))
+        pert_dets = rng.standard_normal((9, 3, 7))  # ONE LESS than pert mask mat.
+        pert_mask = rng.integers(
             low=0, high=2, size=(10, 15, 25), dtype='int')
         with pytest.raises(
             ValueError,
@@ -68,9 +70,10 @@ def test_detections_classes_mismatch(self) -> None:
         Test mismatch in number of classes between perturbed and reference detections.
         """
         impl = DRISEScoring()
-        ref_dets = np.random.rand(2, 8)  # ONE MORE than pert dets mat.
-        pert_dets = np.random.rand(10, 3, 7)
-        pert_mask = np.random.randint(
+        rng = np.random.default_rng(seed=0)
+        ref_dets = rng.standard_normal((2, 8))  # ONE MORE than pert dets mat.
+        pert_dets = rng.standard_normal((10, 3, 7))
+        pert_mask = rng.integers(
             low=0, high=2, size=(10, 15, 25), dtype='int')
         with pytest.raises(
             ValueError,
diff --git a/tests/impls/gen_image_classifier_blackbox_sal/test_rise.py b/tests/impls/gen_image_classifier_blackbox_sal/test_rise.py
index 9d766f0a..00167a06 100644
--- a/tests/impls/gen_image_classifier_blackbox_sal/test_rise.py
+++ b/tests/impls/gen_image_classifier_blackbox_sal/test_rise.py
@@ -35,10 +35,10 @@ def test_configuration(self) -> None:
             assert isinstance(inst_g, RISEScoring)
             assert inst_p.n == 444
             assert inst_p.s == 33
-            assert inst_p.p1 == .22
+            assert np.allclose(inst_p.p1, .22)
             assert inst_p.seed == 42
             assert inst_p.threads == 99
-            assert inst_g.p1 == .22
+            assert np.allclose(inst_g.p1, .22)
             assert inst_i.get_config()['debiased'] is True
 
     def test_configuration_not_debiased(self) -> None:
@@ -58,8 +58,8 @@ def test_configuration_not_debiased(self) -> None:
             inst_g = inst_i._po._generator
             assert isinstance(inst_p, RISEGrid)
             assert isinstance(inst_g, RISEScoring)
-            assert inst_p.p1 == .22
-            assert inst_g.p1 == 0.0
+            assert np.allclose(inst_p.p1, .22)
+            assert np.allclose(inst_g.p1, 0.0)
             assert inst_i.get_config()['debiased'] is False
 
     def test_generation_rgb(self) -> None:
diff --git a/tests/impls/gen_object_detector_blackbox_sal/test_drise.py b/tests/impls/gen_object_detector_blackbox_sal/test_drise.py
index f69e68b2..48cbc071 100644
--- a/tests/impls/gen_object_detector_blackbox_sal/test_drise.py
+++ b/tests/impls/gen_object_detector_blackbox_sal/test_drise.py
@@ -37,7 +37,7 @@ def test_configuration(self) -> None:
             assert isinstance(inst_g, DRISEScoring)
             assert inst_p.n == 123
             assert inst_p.s == 8
-            assert inst_p.p1 == 0.73
+            assert np.allclose(inst_p.p1, .73)
             assert inst_p.seed == 98
             assert inst_p.threads == 5
             assert inst_g.get_config() == {}
diff --git a/tests/impls/gen_object_detector_blackbox_sal/test_random_grid.py b/tests/impls/gen_object_detector_blackbox_sal/test_random_grid.py
index 23f56c2d..f0f3294e 100644
--- a/tests/impls/gen_object_detector_blackbox_sal/test_random_grid.py
+++ b/tests/impls/gen_object_detector_blackbox_sal/test_random_grid.py
@@ -38,7 +38,7 @@ def test_configuration(self) -> None:
             assert isinstance(inst_g, DRISEScoring)
             assert inst_p.n == 55
             assert inst_p.s == (15, 8)
-            assert inst_p.p1 == 0.34
+            assert np.allclose(inst_p.p1, 0.34)
             assert inst_p.seed == 7
             assert inst_p.threads == 2
             assert inst_g.get_config() == {}
diff --git a/tests/impls/perturb_image/test_random_grid.py b/tests/impls/perturb_image/test_random_grid.py
index f9f34244..49c509a5 100644
--- a/tests/impls/perturb_image/test_random_grid.py
+++ b/tests/impls/perturb_image/test_random_grid.py
@@ -21,7 +21,7 @@ def test_init_valued(self) -> None:
 
         assert impl.n == test_n
         assert impl.s == test_s
-        assert impl.p1 == test_p1
+        assert np.allclose(impl.p1, test_p1)
         assert impl.seed == test_seed
         assert impl.threads == test_threads
 
@@ -41,7 +41,7 @@ def test_standard_config(self) -> None:
         for inst in configuration_test_helper(impl):
             assert inst.n == test_n
             assert inst.s == test_s
-            assert inst.p1 == test_p1
+            assert np.allclose(impl.p1, test_p1)
             assert inst.seed == test_seed
             assert inst.threads == test_threads
 
@@ -53,7 +53,8 @@ def test_if_random(self) -> None:
         impl1 = RandomGrid(n=3, s=(5, 4), p1=0.5)
         impl2 = RandomGrid(n=3, s=(5, 4), p1=0.5)
 
-        img = np.random.randint(0, 255, size=(20, 20), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(20, 20), dtype=np.uint8)
 
         masks1 = impl1(img)
         masks2 = impl2(img)
@@ -68,7 +69,8 @@ def test_seed(self) -> None:
         impl1 = RandomGrid(n=3, s=(2, 1), p1=0.6, seed=5)
         impl2 = RandomGrid(n=3, s=(2, 1), p1=0.6, seed=5)
 
-        img = np.random.randint(0, 255, size=(10, 10), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(10, 10), dtype=np.uint8)
 
         masks1 = impl1(img)
         masks2 = impl2(img)
@@ -80,8 +82,8 @@ def test_call_idempotency(self) -> None:
         Test that perturbation generation is idempotent, at least when seeded
         and single-threaded.
         """
-
-        img = np.random.randint(0, 255, size=(10, 10), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(10, 10), dtype=np.uint8)
 
         impl = RandomGrid(n=2, s=(1, 2), p1=0.4, seed=1, threads=0)
 
@@ -95,8 +97,8 @@ def test_perturb_1_channel(self) -> None:
         Test mask generation on a one-channel image of a known size. Number
         of channels should not affect output masks.
         """
-
-        img = np.random.randint(0, 255, size=(4, 6), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(4, 6), dtype=np.uint8)
 
         impl = RandomGrid(n=2, s=(2, 2), p1=0.5, seed=123, threads=0)
 
@@ -112,8 +114,8 @@ def test_perturb_3_channel(self) -> None:
         Test mask generation on a three-channel image of a known size. Number
         of channels should not affect output masks.
         """
-
-        img = np.random.randint(0, 255, size=(4, 6, 3), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(4, 6, 3), dtype=np.uint8)
 
         impl = RandomGrid(n=2, s=(2, 2), p1=0.5, seed=123, threads=0)
 
@@ -129,11 +131,11 @@ def test_multiple_image_size(self) -> None:
         Test that a single implementation can be used for images of varying
         sizes.
         """
-
         impl = RandomGrid(n=5, s=(3, 4), p1=0.2, seed=42, threads=0)
 
-        img_small = np.random.randint(0, 255, size=(5, 7), dtype=np.uint8)
-        img_large = np.random.randint(0, 255, size=(55, 77), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img_small = rng.integers(0, 255, size=(5, 7), dtype=np.uint8)
+        img_large = rng.integers(0, 255, size=(55, 77), dtype=np.uint8)
 
         masks_small = impl(img_small)
         masks_large = impl(img_large)
@@ -148,8 +150,8 @@ def test_threading(self) -> None:
         """
         Test that using threading does not affect results.
         """
-
-        img = np.random.randint(0, 255, size=(4, 6), dtype=np.uint8)
+        rng = np.random.default_rng(seed=0)
+        img = rng.integers(0, 255, size=(4, 6), dtype=np.uint8)
 
         impl = RandomGrid(n=2, s=(2, 2), p1=0.5, seed=123, threads=1)
 
diff --git a/tests/utils/test_detection.py b/tests/utils/test_detection.py
index 5eadbb82..c9e49913 100644
--- a/tests/utils/test_detection.py
+++ b/tests/utils/test_detection.py
@@ -11,8 +11,9 @@ def test_default_objectness_fill(self) -> None:
         Test that when objectness scores are not provided the expected default
         is filled into the appropriate column.
         """
-        test_bbox_mat = np.random.randint(0, 255, (16, 4))
-        test_class_mat = np.random.rand(16, 8)
+        rng = np.random.default_rng(seed=0)
+        test_bbox_mat = rng.integers(0, 255, (16, 4))
+        test_class_mat = rng.standard_normal((16, 8))
         combined_mat = format_detection(test_bbox_mat, test_class_mat)
         # The objectness scores should be in the index 4 column and they should
         # all be 1's.
@@ -22,8 +23,9 @@ def test_explicit_objectness(self) -> None:
         """
         Test that input objectness vector is represented in the output.
         """
-        test_bbox_mat = np.random.randint(0, 255, (16, 4))
-        test_class_mat = np.random.rand(16, 8)
+        rng = np.random.default_rng(seed=0)
+        test_bbox_mat = rng.integers(0, 255, (16, 4))
+        test_class_mat = rng.standard_normal((16, 8))
         test_obj_v = np.tile([.1, .2, .3, .4], 4)
         combined_mat = format_detection(test_bbox_mat, test_class_mat, test_obj_v)
         # that the output objectness column is equivalent to the input.
@@ -34,8 +36,9 @@ def test_explicit_objectness_2d(self) -> None:
         Test that an input objectness vector that is of shape `[nDets x 1]` is
         treated fine.
         """
-        test_bbox_mat = np.random.randint(0, 255, (16, 4))
-        test_class_mat = np.random.rand(16, 8)
+        rng = np.random.default_rng(seed=0)
+        test_bbox_mat = rng.integers(0, 255, (16, 4))
+        test_class_mat = rng.standard_normal((16, 8))
         test_obj_v = np.tile([.1, .2, .3, .4], 4).reshape(16, 1)
         combined_mat = format_detection(test_bbox_mat, test_class_mat, test_obj_v)
         # Test that the output objectness column is equivalent to the input.
@@ -73,8 +76,9 @@ def test_bbox_class_shape_mismatch(self) -> None:
         mismatch in input bbox and classification matrices.
         """
         # 16 boxes, 14 classifications.
-        test_bbox_mat = np.random.randint(0, 255, (16, 4))
-        test_class_mat = np.random.rand(14, 8)
+        rng = np.random.default_rng(seed=0)
+        test_bbox_mat = rng.integers(0, 255, (16, 4))
+        test_class_mat = rng.standard_normal((14, 8))
         with pytest.raises(
             ValueError,
             match=r"along dimension 0, the array at index 0 has size 16 and "
@@ -87,9 +91,10 @@ def test_objectness_shape_mismatch(self) -> None:
         Test that an error is raised when the explicitly input objectness array
         is not a matching size.
         """
-        test_bbox_mat = np.random.randint(0, 255, (16, 4))
-        test_class_mat = np.random.rand(16, 8)
-        test_objnes_v = np.random.rand(11)
+        rng = np.random.default_rng(seed=0)
+        test_bbox_mat = rng.integers(0, 255, (16, 4))
+        test_class_mat = rng.standard_normal((16, 8))
+        test_objnes_v = rng.standard_normal((11))
         with pytest.raises(
             ValueError,
             match=r"along dimension 0, the array at index 0 has size 16 and "
@@ -111,8 +116,9 @@ def test_against_type_upcasting(
         Test that the output is not of a type that is larger than is input.
         E.g. when input is float32, output is *not* float64, but still float32.
         """
-        bbox_mat = np.random.randn(100, 4).astype(bbox_type)
-        classification_mat = np.random.randn(100, 10).astype(clf_type)
+        rng = np.random.default_rng(seed=0)
+        bbox_mat = rng.standard_normal((100, 4)).astype(bbox_type)
+        classification_mat = rng.standard_normal((100, 10)).astype(clf_type)
         output = format_detection(bbox_mat, classification_mat)
         assert output.dtype == expected_type
 
@@ -132,8 +138,9 @@ def test_against_type_upcasting_obj(
         Same as ``test_against_type_upcasting`` but including the objectness
         input vector instead of it being ``None``.
         """
-        bbox_mat = np.random.randn(100, 4).astype(bbox_type)
-        classification_mat = np.random.randn(100, 10).astype(clf_type)
-        objectness = np.random.randn(100).astype(obj_type)
+        rng = np.random.default_rng(seed=0)
+        bbox_mat = rng.standard_normal((100, 4)).astype(bbox_type)
+        classification_mat = rng.standard_normal((100, 10)).astype(clf_type)
+        objectness = rng.standard_normal(100).astype(obj_type)
         output = format_detection(bbox_mat, classification_mat, objectness)
         assert output.dtype == expected_type
diff --git a/tests/utils/test_masking.py b/tests/utils/test_masking.py
index ebb86859..bc125b84 100644
--- a/tests/utils/test_masking.py
+++ b/tests/utils/test_masking.py
@@ -467,7 +467,8 @@ def test_against_type_upcasting(
         based on the input data types.
         E.g. when input is float32, output is *not* float64, but still float32.
         """
-        scalar_vec = np.random.randn(100, 10).astype(scalar_type)
+        rng = np.random.default_rng(seed=0)
+        scalar_vec = rng.standard_normal((100, 10)).astype(scalar_type)
         masks = np.ones((100, 224, 224)).astype(mask_type)
         output = weight_regions_by_scalar(scalar_vec, masks, inv_masks, normalize)
         assert output.dtype == expected_output_type
diff --git a/xaitk_saliency/utils/masking.py b/xaitk_saliency/utils/masking.py
index e22c9583..764cd7a5 100644
--- a/xaitk_saliency/utils/masking.py
+++ b/xaitk_saliency/utils/masking.py
@@ -227,7 +227,8 @@ def benchmark_occlude_image(
     `[H x W [x C]]`.
     """
     img_mat = np.ones((*img_shape, img_channels), dtype=np.uint8)
-    masks = (np.random.rand(num_masks, *img_shape[:2]) < 0.5)
+    rng = np.random.default_rng(seed=0)
+    masks = (rng.standard_normal((num_masks, *img_shape[:2])) < 0.5)
     fill_1c: int = 0
     fill_mc = [0] * img_channels
     perf_counter = time.perf_counter